withDb

Executes block once against the admitted current DB instance, bounding the caller's wait at withDbTimeoutMillis.

A callback is never replayed after it starts: an arbitrary block can perform one side effect and then fail, so transparently invoking it again against another pool could duplicate or split a logical write. Pool-timeout recovery may reopen the active database for future calls, but the original failure is still propagated.

Containment is per pool, not process-wide: the callback runs on the pool's containment lane (poolLanes) in a manager-scoped child coroutine that owns the writer registration and releases it in its own finally. Only the wait is bounded — a started block still runs to completion under NonCancellable, so a bounded DB-critical section is never torn apart mid-write. A block that never returns (Room 3.x logs and retries connection-pool acquisition instead of throwing, so a leaked permit hangs forever) is abandoned rather than cancelled: the caller fails with DatabaseOperationTimeoutException and the active pool is reopened, so later calls are admitted against the replacement pool and its fresh lane instead of queueing behind the wedge.

Long-lived Flow/Paging reads must stay out of withDb; see observeCurrentDb and withReadDb.