On Mon, 20 Feb 2023 19:38:35 GMT, Daniel Jeliński <[email protected]> wrote:
> Please review this fix for a race in KeepAliveCache. > > The sweeper thread (`KeepAliveCache.run()`) terminates when the cache is > empty. The check for empty cache was performed without holding the cache > lock. As a result, there was a small window of time where a new connection > could be added to the cache while the sweeper thread was stopping. The added > connection would then be removed from cache without closing when a new > sweeper thread was started. > > As an additional observation, the check for empty cache was performed after > sweeping. The cache could be empty because all connections were closed, or > because all connections were busy. In the latter case, a new thread was > created soon after the old one terminated. > > This patch addresses both these issues. The sweeper thread makes the decision > to terminate while holding the lock, and other threads are aware of that when > they acquire the lock. Also, the sweeper thread only terminates if the cache > is empty AND there was no cache activity in the last `LIFETIME` (5 seconds). > > Tier1-3 clean. No new tests, the issue was very hard to reproduce without > adding delays in production code.. Looks reasonable to me. Good analysis. But now that the `cacheLock` is used everywhere, do we still need the other lock in the `ClientVector`? It seems that all accesses to the client vector are already protected by the coarser `cacheLock`. ------------- Marked as reviewed by dfuchs (Reviewer). PR: https://git.openjdk.org/jdk/pull/12676
