Marko Mäkelä added a comment - 2024-06-26 05:08 The parameter innodb_doublewrite was made settable at runtime in MDEV-33545 . Why would you want to revert that? The crash occurs due to a missing condition related to updating the debug-only variable innodb_buf_flush_list_now . That we could fix easily: diff --git a/storage/innobase/handler/ha_innodb.cc b/storage/innobase/handler/ha_innodb.cc index cd460b20c1a..a8acf2e4f2f 100644 --- a/storage/innobase/handler/ha_innodb.cc +++ b/storage/innobase/handler/ha_innodb.cc @@ -18408,7 +18408,7 @@ static void buf_flush_list_now_set(THD*, st_mysql_sys_var*, void*, const void* save) { - if (!*(my_bool*) save) + if (!*(my_bool*) save || high_level_read_only) return; const uint s= srv_fil_make_page_dirty_debug; mysql_mutex_unlock(&LOCK_global_system_variables); The above patch is for 10.5 (the earliest supported version). But, 10.5 does not crash because of the following: void buf_dblwr_t::flush_buffered_writes() { if (!is_initialised() || !srv_use_doublewrite_buf) { fil_flush_file_spaces(); return ; } and the following in srv_start() ( innodb_force_recovery=6 will also imply innodb_read_only=ON ): if (srv_force_recovery) { ib::info() << "!!! innodb_force_recovery is set to " << srv_force_recovery << " !!!" ; if (srv_force_recovery == SRV_FORCE_NO_LOG_REDO) { srv_read_only_mode = true ; } }   if (srv_read_only_mode) { sql_print_information( "InnoDB: Started in read only mode" ); srv_use_doublewrite_buf = false ; } In MDEV-33545 , the last part was changed to the following: diff --git a/storage/innobase/srv/srv0start.cc b/storage/innobase/srv/srv0start.cc index 3c90186374c..875dbbe8d57 100644 --- a/storage/innobase/srv/srv0start.cc +++ b/storage/innobase/srv/srv0start.cc @@ -1193,7 +1193,7 @@ dberr_t srv_start( bool create_new_db) if (srv_read_only_mode) { sql_print_information( "InnoDB: Started in read only mode" ); - srv_use_doublewrite_buf = false ; + buf_dblwr.use = buf_dblwr.USE_NO; } high_level_read_only = srv_read_only_mode I will apply the above patch fix 10.5, even though I can confirm that this regression was introduced only in 11.0.6. The setting innodb_flush_list_now only exists in debug instrumented builds for the purpose of testing that the doublewrite buffer can heal some corruption.

Marko Mäkelä added a comment - 2024-06-26 05:17 I realized that innodb_buf_flush_list_now is not the actual culprit, but the SET GLOBAL innodb_doublewrite is. The following fixes this: diff --git a/storage/innobase/handler/ha_innodb.cc b/storage/innobase/handler/ha_innodb.cc index f83120f4076..2723a3890fb 100644 --- a/storage/innobase/handler/ha_innodb.cc +++ b/storage/innobase/handler/ha_innodb.cc @@ -18400,7 +18400,8 @@ static void innodb_data_file_write_through_update(THD *, st_mysql_sys_var*, static void innodb_doublewrite_update(THD *, st_mysql_sys_var*, void *, const void *save) { - fil_system.set_use_doublewrite(*static_cast<const ulong*>(save)); + if (!srv_read_only_mode) + fil_system.set_use_doublewrite(*static_cast<const ulong*>(save)); } static void innodb_log_file_size_update(THD *thd, st_mysql_sys_var*, Here is a little better test case (not using the dangerous setting innodb_force_recovery ): --source include/have_innodb.inc --source include/have_debug.inc --let $restart_parameters=--innodb-read-only --source include/restart_mysqld.inc SET GLOBAL innodb_doublewrite=1,innodb_buf_flush_list_now=1;

Ramesh Sivaraman added a comment - 2024-06-26 05:22 - edited I did not notice MDEV-33545 . innodb_doublewrite was read-only in my local builds (on older versions), and I haven't seen any changes in the documentation. So I thought it was a bug. Thank you for the better MTR test case https://mariadb.com/kb/en/innodb-system-variables/#innodb_doublewrite

Debarun Banerjee added a comment - 2024-06-26 07:27 Thanks for the patch. Looks good to me.