Lock-free fast path for buffer Pin/Unpin under concurrent readers

[krleonid]

Summary

• Lock-free Pin/Unpin fast path: When a block is already loaded and has active readers (readers > 0), atomically increment/decrement readers via compare-and-swap without acquiring the per-block mutex. Eliminates mutex contention for the common case of hot blocks with multiple concurrent readers.
• shared_mutex for BlockManager block registry: Replace exclusive mutex on the block registry (blocks_lock) with std::shared_mutex. TryGetBlock and BlockIsRegistered now acquire a shared (read) lock, allowing concurrent scans to look up block handles in parallel. RegisterBlock uses a fast shared-lock check before falling back to exclusive lock for insertion.

Benchmark results

Lock-free Pin/Unpin (single table, SELECT * WHERE itemId IN <1000 ids>, 95K rows, 9284 segments)

Connections	Before	After	Improvement
Single	~95ms	~92ms	no regression
10 concurrent	562ms	193ms	66% faster
20 concurrent	1177ms	497ms	58% faster

shared_mutex on blocks_lock (multi-DB, 194-column table, SAMPLE 1000 rows)

Setup	Before	After	Improvement
20 DBs, 20 conns	59.3 qps / 334ms	66.1 qps / 300ms	10% faster
40 DBs, 40 conns	56.8 qps / 699ms	60.7 qps / 655ms	6% faster
10 DBs, 16 conns	61.7 qps / 259ms	68.5 qps / 233ms	10% faster

Test plan

• Verify single-connection queries show no regression
• Verify concurrent wide-table scans show reduced latency
• Verify no crashes under concurrent Pin/Unpin stress
• Verify RegisterBlock/UnregisterBlock still work correctly under concurrent load

🤖 Generated with Claude Code