4 Commits

Author SHA1 Message Date
xfy
e44cfc7128 perf(handler): eliminate read-lock upgrade in FileInfoCache.Get with approximate LRU
FileInfoCache.Get previously acquired TWO locks on every cache hit:
1. RLock → check existence + TTL
2. Release RLock → Lock → MoveToFront (LRU update) → Unlock

The MoveToFront on every hit forced a read-to-write lock upgrade,
creating contention under concurrent reads.

Apply approximate LRU: skip MoveToFront on Get (read) path entirely.
LRU position is only updated in Set (write) path. This is the same
pattern already used by internal/cache/file_cache.go.

Result: Get fast path reduced from 2 lock acquisitions to 1 RLock.
TTL-expired entries still use double-check locking for safe removal.
2026-06-04 00:13:29 +08:00
xfy
2734b04d8f refactor: remove 16.8k lines of dead code across all internal packages
- Delete unused files: tempfile subsystem, matcher variants, server/internal
- Remove 200+ unused functions across proxy, ssl, lua, http2/3, stream, variable
- Fix proxy test type errors (backgroundRefresh ctx→Request)
- Move bench/tools mock backend into internal/testutil
- Remove corresponding test functions for all deleted code
2026-06-03 16:15:43 +08:00
xfy
5f470993ff perf(handler): use RWMutex for FileInfoCache
- Change FileInfoCache.mu from sync.Mutex to sync.RWMutex
- Get method uses RLock for concurrent read access
- Stats method uses RLock for read-only operation
- Double-check pattern for lock upgrades (TTL expiry, LRU move)

This improves concurrent read performance for the FileInfo cache,
which is read-heavy in static file serving scenarios.

Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
2026-04-30 16:15:39 +08:00
xfy
8cc3fdef6f perf(handler): optimize static file serving performance
- Add pathPrefixLen for zero-allocation path stripping
- Precompute ETag in FileCache.Set, reuse on cache hits
- Add MIME LRU cache with O(1) operations using container/list
- Remove sharded cache (eager LRU was slower than single-lock)
- Add FileInfo cache to reduce os.Stat calls (TTL-only strategy)
- Adjust test expectations for normalized root path format

Benchmark results: Lookup 12.7µs → 10.6µs (16% improvement)

Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
2026-04-30 14:17:56 +08:00