Symptom

TSan reports data races in test/csuite/schema_disagg_abort itself (not in WiredTiger) in the two-node lone-follower runs of smoke.sh (-r lf ...). Seen in PR #14814's patch 6ab9dfe7c9593e0007815e7e on amazon2023-arm64-tsan (-r lf -k l8), and locally on a macOS arm64 TSan build in -r lf and -r lf -k l8.

WARNING: ThreadSanitizer: data race
  Write of size 1 by thread T24:
    #0 thread_reader_run reader.c:79
  Previous read of size 1 by main thread:
    #0 node_trigger_wait node.c:171
    #1 node_run node.c:398

WARNING: ThreadSanitizer: data race
  Write of size 4 by thread T22:
    #0 pipe_relay_event event_pipe.c:41
    #1 apply_event worker.c:293
  Previous read of size 4 by thread T21:
    #0 pipe_relay_event event_pipe.c:33
    #1 apply_event worker.c:293

Cause

TEST_CONFIG.peer_alive and TEST_CONFIG.pipe_write_fd are shared by the node's threads without synchronization.

• peer_alive is cleared by a worker when a relay write fails (event_pipe.c:42) and by the reader on pipe EOF (reader.c:79). The generator (generator.c:51, :257, :354), the workers (worker.c:147) and the main thread (node.c:171) read it with plain loads.
• pipe_write_fd: a leader's workers relay concurrently (worker.c:272, :293, :316). The first to see EPIPE closes the descriptor and stores -1 (event_pipe.c:40-41) while the others load it to write (event_pipe.c:33, :36).

Inference, not observed: beyond the TSan report, two workers can both close the same descriptor number, or write an event into a descriptor number the process has meanwhile reused for another file.

Reproduce

Build with TSan and run bash ./smoke.sh from build/test/csuite/schema_disagg_abort. The connection API race (WT-18743) fires first in the single-node runs unless it is fixed or suppressed; these races show up in the -r lf runs.

Impact

Blocks re-enabling the schema_disagg_abort TSan smoke (FIXME-WT-18743 entries in test/evergreen.yml): with WT-18743 fixed, these races still fail it.

Proposed fix

Load and store peer_alive atomically, and keep the out-pipe open for the process's life instead of closing it from a worker, as the node already does with its in-pipe and self-pipe.