For some series of patches that has gone in, the node id is not being
generated or stored in the boundto address properly. This may have
always been a bug or a bug resulting in recent changes. I'm not certain
either way.
The symptoms of the problem are that "message_source_is_local" will
return the wrong value for some messages during race conditions of the
netif_determine function running and updating the nodeid at the same
time as totempg_nodeid_get retrieving the information. Sometime during
this operation, the node id is set to zero, but later is compared
against a proper node value, rejecting messages in many services such as
ckpt, synchronization, and other badness.
The attached patch fixes the problem.
Regards
-steve
Index: exec/totemnet.c
===================================================================
--- exec/totemnet.c (revision 1676)
+++ exec/totemnet.c (working copy)
@@ -692,24 +692,9 @@
int res;
res = totemip_iface_check (bindnet, bound_to,
- interface_up, interface_num);
+ interface_up, interface_num,
+ instance->totem_config->clear_node_high_bit);
- /*
- * If the desired binding is to an IPV4 network and nodeid isn't
- * specified, retrieve the node id from this_ip network address
- *
- * IPV6 networks must have a node ID specified since the node id
- * field is only 32 bits.
- */
- if (bound_to->family == AF_INET && bound_to->nodeid == 0) {
- int32_t nodeid = 0;
- memcpy (&nodeid, bound_to->addr, sizeof (int));
- if(nodeid < 0 && instance->totem_config->clear_node_high_bit) {
- nodeid = 0 - nodeid;
- }
- bound_to->nodeid = nodeid;
- }
-
return (res);
}
Index: exec/totemip.c
===================================================================
--- exec/totemip.c (revision 1676)
+++ exec/totemip.c (working copy)
@@ -420,7 +420,8 @@
int totemip_iface_check(struct totem_ip_address *bindnet,
struct totem_ip_address *boundto,
int *interface_up,
- int *interface_num)
+ int *interface_num,
+ int mask_high_bit)
{
int fd;
struct {
@@ -558,6 +559,18 @@
}
}
finished:
+ /*
+ * Mask 32nd bit off to workaround bugs in other poeples code
+ * if configuration requests it.
+ */
+ if (ipaddr.family == AF_INET && ipaddr.nodeid == 0) {
+ unsigned int nodeid = 0;
+ memcpy (&nodeid, ipaddr.addr, sizeof (int));
+ if ((nodeid & 0x7FFFFFFF) && mask_high_bit) {
+ nodeid &= 0x7FFFFFFF;
+ }
+ ipaddr.nodeid = nodeid;
+ }
totemip_copy (boundto, &ipaddr);
close(fd);
return 0;
Index: exec/totemip.h
===================================================================
--- exec/totemip.h (revision 1676)
+++ exec/totemip.h (working copy)
@@ -71,7 +71,7 @@
extern int totemip_totemip_to_sockaddr_convert(struct totem_ip_address *ip_addr,
uint16_t port, struct sockaddr_storage *saddr, int *addrlen);
extern int totemip_parse(struct totem_ip_address *totemip, char *addr, int family);
-extern int totemip_iface_check(struct totem_ip_address *bindnet, struct totem_ip_address *boundto, int *interface_up, int *interface_num);
+extern int totemip_iface_check(struct totem_ip_address *bindnet, struct totem_ip_address *boundto, int *interface_up, int *interface_num, int mask_high_bit);
/* These two simulate a zero in_addr by clearing the family field */
static inline void totemip_zero_set(struct totem_ip_address *addr)
_______________________________________________
Openais mailing list
[email protected]
https://lists.linux-foundation.org/mailman/listinfo/openais
