Kristian Nielsen added a comment - 2023-11-08 12:22 - edited The patch moves the check of trx_is_interrupted() to before taking the lock_sys.wait_mutex.m_mutex. But checking for kill always needs to be done while holding the mutex that's being used in pthread_cond_wait: err= my_cond_timedwait(&trx->lock.cond, &lock_sys.wait_mutex.m_mutex, &abstime); Otherwise the kill may occur just after the check and be ignored, and the kill is lost. This will break parallel replication, as deadlocks remain unhandled and the slave will hang. It sems necessary to be able to call thd_kill_level() while holding other mutex, as this is a requirement for correctly handling kill. I understand the convenience of trying to run pending apc from thd_kill_level(), since it is something we can expect to be called frequently. But it doesn't seem safe. What about using mysql_mutex_trylock() in thd_kill_level(), and only running the apc if the lock can be obtained? The request can then be dequeued under LOCK_thd_kill. Then the apc should be run while temporarily unlocking LOCK_thd_kill. Update: No, the lock cannot be released while running the apc, it protects the lifetime of the request. But that might be ok, just a retriction on which mutexes can be taken in an apc.

Kristian Nielsen added a comment - 2023-11-08 12:57 Here an RFC patch for the above idea: https://github.com/MariaDB/server/commit/d35f30214a2a1da0f5dbc6e69d1bd9a3a5e98b06

Sergei Golubchik added a comment - 2023-11-08 13:54 Thanks! Looks good, I'll apply and push it, if you don't mind

Kristian Nielsen added a comment - 2023-11-08 14:06 Sure, please do. Meanwhile, I'll see if I can come up with a quick testcase, depends on how tricky it is to get the threads coordinated.

Kristian Nielsen added a comment - 2023-11-08 17:24 Pushed a testcase to knielsen_mdev32728: https://github.com/MariaDB/server/commit/2d0eb0ddfc13d7b5a788450a281a3c5b5854482e