ben-manes / caffeine / #5552

100%

master: 100%

Build Type

push

github

Committed by ben-manes

Commit Message

discard an in-flight refresh on the put insert path

The put() insert was the only write path that didn't clear an in-flight
refresh token, so an explicit refresh() on an absent key could linger in
`refreshes` after a concurrent insert and suppress later refreshes until
it completed. The insert now discards inside data.computeIfAbsent's
lambda -- under the bin lock, no node monitor -- so it wins like the
update and remove paths. The overwrite the audit reported was a false
positive: the refresh completion guards already reject a stale reload, so
this is a consistency fix, not a correctness one.

The discard must run inside the CHM lambda, never under synchronized(node)
across putIfAbsent: putIfAbsent locks bin heads internally for both the
link and a follow-on resize (addCount -> transfer, after the node is
visible), so holding the node monitor across it is a node-before-bin
acquisition that deadlocks a concurrent computeIfPresent remove. The
synchronized variant shipped and was reverted after MultiThreadedTest
deadlocked; documented in synchronization.md.

Also re-check node.isAlive() in refreshIfNeeded's registration lambda so
a node retired between the gate snapshot and the lambda doesn't launch a
doomed reload.

Tests: RefreshAfterWriteTest.put_insert_discardsRefresh,
BoundedLocalCacheTest.refreshIfNeeded_retiredAfterSnapshot. Also folds in
minor dependency and build/CI bumps (dependency-analysis, javapoet, etc.).

Coverage Stats

4050 of 4056 branches covered (99.85%)

10 of 10 new or added lines in 1 file covered. (100.0%)

1 existing line in 1 file now uncovered.

8244 of 8245 relevant lines covered (99.99%)

1.0 hits per line