In rare circumstances under overload conditions, disconnecting an IPC
connection from IPC using the checkpoint service and thousands of
checkpoints can result in an assertion in the executive.
The solution is that the close routine should return -1 to the
lib_exit_fn function, which should then return and continue processing
later as per the API semantics.
Patch attached to fix.
Regards
-steve
Index: exec/ckpt.c
===================================================================
--- exec/ckpt.c (revision 1641)
+++ exec/ckpt.c (working copy)
@@ -975,6 +975,7 @@
struct list_head *list;
struct checkpoint_section *section;
+printf ("releasing checkpoint %s\n", checkpoint->name.value);
openais_timer_delete (checkpoint->retention_timer);
/*
@@ -1017,9 +1018,7 @@
iovec.iov_base = (char *)&req_exec_ckpt_checkpointclose;
iovec.iov_len = sizeof (req_exec_ckpt_checkpointclose);
- assert (totempg_groups_mcast_joined (openais_group_handle, &iovec, 1, TOTEMPG_AGREED) == 0);
-
- return (-1);
+ return (totempg_groups_mcast_joined (openais_group_handle, &iovec, 1, TOTEMPG_AGREED));
}
static int ckpt_exec_init_fn (struct objdb_iface_ver0 *objdb)
@@ -2357,6 +2356,7 @@
struct checkpoint_cleanup *checkpoint_cleanup;
struct list_head *list;
struct ckpt_pd *ckpt_pd = (struct ckpt_pd *)openais_conn_private_data_get (conn);
+ unsigned int res;
log_printf (LOG_LEVEL_DEBUG, "checkpoint exit conn %p\n", conn);
@@ -2370,9 +2370,16 @@
struct checkpoint_cleanup, list);
assert (checkpoint_cleanup->checkpoint_name.length != 0);
- ckpt_checkpoint_close (
+ res = ckpt_checkpoint_close (
&checkpoint_cleanup->checkpoint_name,
checkpoint_cleanup->ckpt_id);
+ /*
+ * If checkpoint_close fails because of full totem queue
+ * return -1 ande try again later
+ */
+ if (res == -1) {
+ return (-1);
+ }
list_del (&checkpoint_cleanup->list);
free (checkpoint_cleanup);
_______________________________________________
Openais mailing list
[email protected]
https://lists.linux-foundation.org/mailman/listinfo/openais
