Evergreen patch:EVG Patch

  Affected Tests (10 of 57):
  1. unsupported_read_write_concerns.js
  2. mongod_waits_for_cms.js
  3. read_concern_target_time_wait.js
  4. simple_replica_set.js
  5. simple_one_node_restart.js
  6. change_stream_can_read_from_secondary.js
  7. large_oplog_batch_application.js
  8. establish_connection_during_stepup.js
  9. transaction_table_oplog_replay.js
  10. replSetInitiate_no_config.js

  Details:

  During StorageEngineImpl::cleanShutdown(), the spill WiredTiger engine's shutdown path closes the WiredTiger connection via wiredtiger_close(). Inside that call, __wti_prefetch_destroy() tears down the prefetch thread group by calling __wt_thread_group_destroy() →
  {}thread_group_shrink() → {}wt_session_close_internal(), which performs a memset (write) over a WiredTiger session struct in-place. Concurrently, WiredTiger's internal sweep server thread ({_}_sweep_server) is still running and calling __wt_session_array_walk(), which
  performs an atomic read of the same session struct memory — the two accesses race.

  Thread A — shutdown:
    memset ← WRITE (8 bytes)
    __wt_session_close_internal
    __thread_group_shrink
    __wt_thread_group_destroy
    __wti_prefetch_destroy
    __conn_close
    SpillWiredTigerKVEngine::cleanShutdown()
    StorageEngineImpl::cleanShutdown()

  Thread B — sweep server:
    __wt_session_array_walk ← ATOMIC READ (same address)
    __sweep_server

  Root Cause:

  WiredTiger's __conn_close shuts down the prefetch thread group — and closes those threads' sessions — before stopping the sweep server thread. The sweep server's __wt_session_array_walk continues reading the session array during this window, overlapping with the
  memset that zeros out the session struct being freed. This is a WiredTiger-internal shutdown ordering bug: the sweep server should be stopped before any sessions are destroyed.