mirror of
https://github.com/0xERR0R/blocky.git
synced 2026-10-12 02:01:13 +00:00
Max
1900670c18
The wildcard cache stored each rule as trie nodes, about 86 bytes per rule, so HaGeZi TIF alone held 150 MiB of heap. Rules are now grouped by TLD and, within a TLD, sorted and concatenated into one string per base length, the layout the exact cache already uses. A lookup splits off the TLD and checks each suffix of the remainder from the broadest parent down. A hash filter with two bits per rule in one 64-bit word rejects most suffixes before the binary search. The broadest stored parent still wins, a *.tld rule still covers the whole TLD, and the reported rule string is unchanged. TIF drops to 24.5 MiB. A refresh built every group in parallel, so a process with four large lists peaked near 880 MiB RSS. Each ListCache now rebuilds one group at a time and admits one refresh call at a time; downloads inside a group keep blocking.loading.concurrency. A caller whose context ends while waiting for its turn returns ctx.Err() instead of starting another refresh afterwards. The old group stays active until its replacement is ready. Peak RSS through a refresh drops to about 260 MiB and the refresh takes about a third longer. The exact, regex and wildcard caches build the logrus entry for a hit only when debug logging is enabled, which removes the per-hit allocations when it is off.