cel_pgsql: deadlock on unload and core_event_dispatcher

A deadlock can happen between a thread unloading or reloading the cel_pgsql
module and the core_event_dispatcher taskprocessor thread. Description of
what is happening:

Thread 1 (for example, a netconsole thread):

    a "module reload cel_pgsql" is launched
    the thread enter the "my_unload_module" function (cel_pgsql.c)
    the thread acquire the write lock on psql_columns
    the thread enter the "ast_event_unsubscribe" function (event.c)
    the thread try to acquire the write lock on ast_event_subs[sub->type]

Thread 2 (core_event_dispatcher taskprocessor thread):

    the taskprocessor pop a CEL event
    the thread enter the "handle_event" function (event.c)
    the thread acquire the read lock on ast_event_subs[sub->type]
    the thread callback the "pgsql_log" function (cel_pgsql.c), since it's a subscriber of CEL events
    the thread try to acquire a read lock on psql_columns

(closes issue ASTERISK-22854)
Reported by: Etienne Lessard
Patches:
     cel_pgsql_fix_deadlock_event.patch uploaded by hexanol (license 6394)
........

Merged revisions 404603 from http://svn.asterisk.org/svn/asterisk/branches/1.8
........

Merged revisions 404604 from http://svn.asterisk.org/svn/asterisk/branches/11
........

Merged revisions 404605 from http://svn.asterisk.org/svn/asterisk/branches/12

git-svn-id: https://origsvn.digium.com/svn/asterisk/trunk@404606 65c4cc65-6c06-0410-ace0-fbb531ad65f3

diff --git a/cel/cel_pgsql.c b/cel/cel_pgsql.c
index 5ba5b99..8db85d7 100644 (file)
--- a/cel/cel_pgsql.c
+++ b/cel/cel_pgsql.c
@@ -346,8 +346,9 @@ ast_log_cleanup:
 static int my_unload_module(void)
 {
        struct columns *current;
-       AST_RWLIST_WRLOCK(&psql_columns);
+
        ast_cel_backend_unregister(PGSQL_BACKEND_NAME);
+       AST_RWLIST_WRLOCK(&psql_columns);
        if (conn) {
                PQfinish(conn);
                conn = NULL;