handle_fatal_signal (sig=6) in galera::FSM<galera::Replicator::State, galera::ReplicatorSMM::Transition, galera::EmptyGuard, galera::ReplicatorSMM::StateAction>::shift_to | galera/src/fsm.hpp:81

commit 12a9e13401b4d9e7b781034eddf04ffe957c07d0
Author: Krunal Bauskar <krunal.bauskar@percona.com>
Date: Wed Nov 29 17:17:22 2017 +0530

• PXC#806: Early read of query-id result in stale trx handle assignment.

• During an execution of a Stored Procedure (SP) as a background
  action (say through event scheduler), query-id get assigned at
  later stage, before PXC captures it to derive wsrep_next_trx_id.

• PXC uses this wsrep_next_trx_id as unique id to maintain
  TrxHandle and thread mapping.

• This early capture of query-id resulted in assigning an invalid
  wsrep_next_trx_id (read 0). Problem starts if multiple threads
  gets the same invalid (read 0) id as their unique id to look
  for TrxHandle that eventually cause sharing of TrxHandle.

Solution:

--------

• Fixed by ensuring that when PXC reads query_id, it is a valid id.
  If not, then a valid query_id is generated.
  (MySQL flow anyway generates it but at later stage).

• We could have also corrected it by skipping setting of
  wsrep_next_trx_id at the said stage and waiting for MySQL to
  generate it but that means a drill-down code inside
  the main logic. So keeping it at outer layer
  makes it more manageable.

See https://bugs.launchpad.net/percona-xtradb-cluster/+bug/1734312