PT-2534 - fixed connection leak due to confusion on parent flag

[marcelohpf]

Introduces a new argument to Cxn to flag that connection should not be closed automatically with InactiveDestroy, while keep the parent argument for its current usage of topology parent (or connection source parent)

The code had two different purposes to the same argument "parent" - topology parent and process parent in fork scenarios leading to a leak when get_replicas was called multiple times (e.g. when waiting for replication delays).

None of the current percona toolkit bin/ scripts seems to use forking other than deadlock detection. None of the "parent => X" cases were properly used in a multiprocess scenario, even when daemonizing the tool, connections were created after the fork;

You can reproduce it by:

• create large enough table;
• start dev environment mysqls
• trigger OSC:

./bin/pt-online-schema-change \
  'D=mysqlslap,t=t1' --alter "ADD COLUMN mycolumn varchar(255) NOT NULL DEFAULT ''" \
  --execute --no-version-check --statistics --no-check-alter --recurse 2 --max-lag 10 \
  --chunk-size 10 --progress time,30  \
  --recursion-method dsn=D=percona,t=dsns \ # important
  --no-check-replication-filters \
  --database mysqlslap --host 127.0.0.1 \
  --user msandbox --password msandbox --port 12345

Then connect to one of the read-replicas and watch process list grow indefinitely:

while true; do 
  mysql -h 127.0.0.1 -P 12346 -u msandbox -pmsandbox -e 'show processlist;'
  sleep 1;
done

The root cause of the issue is that InactiveDestroy prevent dbh closing connection as part of the DESTROY workflow that is triggered when object looses reference count.

# other code
my $dbh;
{
  $dbh = DBI::connect();
   $dbh -> {InactiveDestroy} = 1;
  fork();
  if (is children) {
     $dbh->ping();  # ok
     exit;
  }
}
$dbh->ping() # should not fail because InactiveDestroy = 1; else it would

• The contributed code is licensed under GPL v2.0
• Contributor Licence Agreement (CLA) is signed
• util/update-modules has been ran
• Documentation updated
• Test suite update

[it-percona-cla]

All committers have signed the CLA.

[EvgeniyPatlan]

I would suggest some changes. They are for test scenario but still it is netter to follow perl recommendations

[EvgeniyPatlan]

LGTM

[marcelohpf]

I had to update the commit author due to the CLA agreement, sorry for the force push;

[EvgeniyPatlan]

thanks you @marcelohpf for the contribution and for the quick fixes!

[charles-001]

I'm pretty sure this is the issue I'm currently running into with the latest version (3.7.1). While pt-osc is running, it's creating a ton of connections on the replica to check its lag (like 1000+). Eventually, it hits max_user_connections and fails. Seems like a regression happened in 3.7.1

Cannot connect to MySQL: DBI connect(';mysql_read_default_file=~/.my.cnf;host=host;mysql_read_default_group=client','',...) failed: User my_user already has more than 'max_user_connections' active connections at /usr/bin/pt-online-schema-change line 2351.