res_pjsip_registrar: lock transport monitor when setting 'removing' flag
authorKevin Harwell <kharwell@digium.com>
Thu, 7 Feb 2019 15:23:37 +0000 (09:23 -0600)
committerKevin Harwell <kharwell@digium.com>
Thu, 7 Feb 2019 20:28:41 +0000 (14:28 -0600)
A previous patch attempt to mitigate blocked threads on transport shutdown for
a given contact. It was thought that a second lock could be avoided by checking
the 'removing' flag on the transport monitor twice (once before and once after
the normal named aor locking). However as with usual threading issues if the
timing was right the original problem still occured.

This patch adds locking around the first 'removing' flag check and set, thus
nullifying the secondary check, so it was removed.

ASTERISK-28213

Change-Id: Iaa8e36e5311789549b76d8de42dfcea96013b2ed

res/res_pjsip_registrar.c

index 83fbdf8..e6b7901 100644 (file)
@@ -341,25 +341,15 @@ static int register_contact_transport_remove_cb(void *data)
 
        aor = ast_sip_location_retrieve_aor(monitor->aor_name);
        if (!aor) {
+               ao2_lock(monitor);
+               monitor->removing = 0;
+               ao2_unlock(monitor);
                ao2_ref(monitor, -1);
                return 0;
        }
 
        ao2_lock(aor);
 
-       /*
-        * We're now locked so check again to make sure some other thread is not
-        * currently removing the contact, or already has.
-        */
-       if (monitor->removing) {
-               ao2_unlock(aor);
-               ao2_ref(aor, -1);
-               ao2_ref(monitor, -1);
-               return 0;
-       }
-
-       monitor->removing = 1;
-
        contact = ast_sip_location_retrieve_contact(monitor->contact_name);
        if (contact) {
                ast_sip_location_delete_contact(contact);
@@ -400,14 +390,15 @@ static void register_contact_transport_shutdown_cb(void *data)
         * same monitor from different threads. Only one of the calls needs to do the
         * actual removing of the contact, so if one is currently removing then any
         * subsequent calls can skip.
-        *
-        * We'll call it non locked here, but check again once locked just in case the
-        * flag was updated (see register_contact_transport_remove_cb).
         */
+       ao2_lock(monitor);
        if (monitor->removing) {
+               ao2_unlock(monitor);
                return;
        }
 
+       monitor->removing = 1;
+
        /*
         * Push off to a default serializer.  This is in case sorcery
         * does database accesses for contacts.  Database accesses may
@@ -416,8 +407,11 @@ static void register_contact_transport_shutdown_cb(void *data)
         */
        ao2_ref(monitor, +1);
        if (ast_sip_push_task(NULL, register_contact_transport_remove_cb, monitor)) {
+               monitor->removing = 0;
                ao2_ref(monitor, -1);
        }
+
+       ao2_unlock(monitor);
 }
 
 AST_VECTOR(excess_contact_vector, struct ast_sip_contact *);
@@ -728,8 +722,8 @@ static void register_aor_core(pjsip_rx_data *rdata,
                                 * the contact won't be valid anymore if that happens.
                                 */
                                contact_name = ast_sorcery_object_get_id(contact);
-                               monitor = ao2_alloc_options(sizeof(*monitor) + 2 + strlen(aor_name)
-                                       + strlen(contact_name), NULL, AO2_ALLOC_OPT_LOCK_NOLOCK);
+                               monitor = ao2_alloc(sizeof(*monitor) + 2 + strlen(aor_name)
+                                       + strlen(contact_name), NULL);
                                if (monitor) {
                                        strcpy(monitor->aor_name, aor_name);/* Safe */
                                        monitor->contact_name = monitor->aor_name + strlen(aor_name) + 1;