David Schilling 65f524d81d MDEV-41450 TOI DDL BF-aborts InnoDB purge, purge loops forever ...

Fix for MDEV-38218 changed MDL conflict handling to allow high priority
cluster threads to BF abort any conflicting MDL lock holder.
With this, the InnoDB purge coordinator became also potential target
for BF aborts.
However, InnoDB purge thread has no statement boundary and could never
reset THD::killed state, after the BF abort. Purge THDs are taken
from a pool of innodb_purge_threads_MAX THDs, so about 32 purge runs
later the killed THD coordinates purge again. From then on every
open_purge_table() for a table with an indexed virtual column fails,
and purge_sys_t::close_and_reopen() retries forever. Purge stops and
server shutdown hangs.

Separately, open_purge_table() allocates a TABLE_LIST on thd->mem_root
and the coordinator THD never freed its mem_root, so it grew with every
purge batch and every close_and_reopen() retry.

Fix:
- wsrep_handle_mdl_conflict(): when the granted thread is a
  SYSTEM_THREAD_GENERIC THD holding explicit MDL, make the BF requestor
  wait instead of aborting it. In practice this covers the InnoDB
  background threads (purge, statistics, FTS optimizer), which release
  their locks on their own. This restores the pre-MDEV-38218 behaviour
  for these threads only.
- open_purge_table(): open with MYSQL_OPEN_IGNORE_KILLED so that a stale
  kill flag cannot block purge.
- trx_purge_close_tables(): call reset_thd() instead of
  close_thread_tables() to free the coordinator's mem_root per batch.

This commit has a new mtr test galera.MDEV-41450

This work is based on pull request https://github.com/MariaDB/server/pull/5854
by David Schilling, which is now rebased to MariaDB 10.11 and extended to
have a mtr test.

Assisted-by: https://mariadb.org/governance/governance-ai-policy/

2026-10-23 09:31:30 +03:00

263 KiB

Raw Permalink Blame History

View Raw