FLUSH TABLES WITH READ LOCK hanging

[Seppo Jaakola]

in wsrep-5.5 the FTWRL command can remain hanging, how to reproduce:
1. start one node
2. issue continuous DML load against the node
3. try a few times to issue FLUSH TABLES WITH READ LOCK

[Seppo Jaakola (seppo-jaakola)]

Observed a related issue with "innodb disallow writes" patch (which is fixed now)

[Seppo Jaakola (seppo-jaakola)]

Change set #3546 pushed. It contains fixes for wsrep MDL integration and lp:836254

FTWRL can still suffer from lock wait timeout. There seems to be unresolved MDL deadlock, which happens also with MySQL native source build.

[Seppo Jaakola (seppo-jaakola)]

changed to fix committed state. MDL deadlock is probably native MySQL issue, it was reproduced with 5.5.12. New bug tracker will be entered to follow that if need be.

[Seppo Jaakola (seppo-jaakola)]

Turns out that FTWRL does still not work correctly, when run in the context of donor thread. Bug re-opened to fix this use case

[Alex Yurchenko (ayurchen)]

It must be noted that the problem with donor thread is that it actually doesn't call flush_tables_with_read_lock() (new function in 5.5) from the donor thread, instead it calls reload_acl_and_cache() (which is the 5.1 way). Consequently
1) donor thread must call flush_tables_with_read_lock()
2) reload_acl_and_cache() was wsrep-modified in 5.1. This modification must be reverted since it is no longer used for FTWRL purpose in 5.5.

[Seppo Jaakola (seppo-jaakola)]

Pushed a fix (#3548) to run FLUSH TABLES WITH READ LOCK command. Tested with rsync SST and 3 nodes join/drop sequences under heavy SQL load.