-
Type:
Bug
-
Resolution: Duplicate
-
Priority:
Major - P3
-
None
-
Affects Version/s: None
-
Component/s: Checkpoints, Sweep Server
-
None
-
Storage Engines - Transactions
-
127.959
-
None
-
None
-
Not Needed
WT-17323 (commit 8334eb7f66) added an atomic read of txn_global.last_ckpt_timestamp in the sweep server (__sweep_close_dhandle_locked, src/conn/conn_sweep.c:172), but all writers still use plain stores (src/checkpoint/checkpoint_txn.c:1887, src/conn/conn_layered_checkpoint_pick_up.c:850). TSan reports the mixed access as a data race.
Fails intermittently in format-stress-test-disagg-switch-tsan (switch mode creates ingest tables for the sweep server to expire). The format run itself succeeds; the process aborts on the TSan report (abort_on_error=1). Reproduced on two unrelated PR patches:
WARNING: ThreadSanitizer: data race (pid=16518)
Write of size 8 at 0xffff8e001120 by thread T107 (mutexes: write M0, write M1):
#0 __checkpoint_db_internal src/checkpoint/checkpoint_txn.c:1887:46
#1 __checkpoint_db_wrapper src/checkpoint/checkpoint_txn.c:2006:11
#2 __wt_checkpoint_db src/checkpoint/checkpoint_txn.c:2085:9
Previous atomic read of size 8 at 0xffff8e001120 by thread T27:
#0 __wt_atomic_load_uint64_relaxed src/include/gcc.h:409:1
#1 __sweep_close_dhandle_locked src/conn/conn_sweep.c:172:13
#2 __sweep_expire_one src/conn/conn_sweep.c:205:5
#3 __sweep_server src/conn/conn_sweep.c:540:13
SUMMARY: ThreadSanitizer: data race src/checkpoint/checkpoint_txn.c:1887:46 in __checkpoint_db_internal
Suggested fix: convert all accesses to last_ckpt_timestamp (declared wt_shared in src/include/txn.h) to atomic loads/stores; same for the adjacent last_ckpt_disaggregated_schema_epoch write.
Related: WT-17323, WT-17263 (PR #14123), WT-17959 (PR #14124).
- is caused by
-
WT-17323 Ingest table cursors are not swept, leading to potential file descriptor exhaustion
-
- Closed
-
- is related to
-
-
- Closed
-
-
-
- Closed
-
-
-
- Closed
-
-
-
- Closed
-
- related to
-
-
- Closed
-