Fix deadlock in pgxpool under concurrent context cancellation

[zhuyh1606-oss]

/claim #1

Problem

Under high-concurrency with rapid context cancellation, pgxpool can deadlock/livelock when:

1. Multiple goroutines call Acquire with contexts that get cancelled mid-flight
2. puddle's semaphore token is consumed but resource init fails due to cancel
3. The constructor goroutine holds the semaphore token until completion
4. pgxpool retry loop doesn't check ctx between iterations

Fix

Two-part fix:

puddle/pool.go acquire()

Added ctx.Done() check after semaphore acquire succeeds but before resource creation. If the context was cancelled between TryAcquire and the check, the semaphore is released immediately instead of starting a doomed constructor goroutine.

pgxpool/pool.go Acquire()

Added ctx.Err() check in the retry loop after Destroy() calls for expired/unhealthy connections. This terminates the retry loop immediately when context is cancelled, instead of cycling through all maxConns+1 attempts.

Testing

Added TestPoolAcquireStressCancelContention - runs 50 goroutines each acquiring with tight context cancellation (5ms timeout), exercising the semaphore contention + cancel path.

[opirebot]

😅 Unfortunately there are no rewards left to claim in this issue!

[zhuyh1606-oss]

Hi @jaasielitaigq, any chance to review this PR? The fix addresses the deadlock in pgxpool under context cancellation with a regression test. Also noticed the Opire reward needs to be + minimum — if you'd like to set it up, I can re-claim the issue.

[zhuyh1606-oss]

Hi @jaasielitaigq, just following up on this PR. It's been open for a while and is ready for merge. Could you please review and merge when you have time? Thanks!