[Cluster-devel] cluster/group/gfs_controld lock_dlm.h recover.c

[teigland@sourceware.org <teigland@sourceware.org>]

CVSROOT:	/cvs/cluster
Module name:	cluster
Changes by:	teigland at sourceware.org	2006-10-23 15:44:34

Modified files:
	group/gfs_controld: lock_dlm.h recover.c 

Log message:
	Patch from Abhi to fix case where a node's mount is rejected by other
	group members causing gfs_controld on the mounter to leave the group
	immediately.  It was sometimes leaving before its join was even
	finished which caused groupd to reject the leave, so we need to wait
	for the join to complete before doing the leave.

Patches:
http://sourceware.org/cgi-bin/cvsweb.cgi/cluster/group/gfs_controld/lock_dlm.h.diff?cvsroot=cluster&r1=1.20&r2=1.21
http://sourceware.org/cgi-bin/cvsweb.cgi/cluster/group/gfs_controld/recover.c.diff?cvsroot=cluster&r1=1.22&r2=1.23

--- cluster/group/gfs_controld/lock_dlm.h	2006/10/16 14:44:02	1.20
+++ cluster/group/gfs_controld/lock_dlm.h	2006/10/23 15:44:33	1.21
@@ -138,6 +138,7 @@
 	int			mount_client_fd;
 	int			mount_client_notified;
 	int			mount_client_delay;
+	int                     group_leave_on_finish;
 	int			remount_client;
 	int			init;
 	int			got_our_options;
--- cluster/group/gfs_controld/recover.c	2006/10/16 17:12:10	1.22
+++ cluster/group/gfs_controld/recover.c	2006/10/23 15:44:33	1.23
@@ -1933,6 +1933,9 @@
 {
 	char buf[MAXLINE];
 	int rv, error = 0;
+	struct mg_member *memb;
+	
+	memb = find_memb_nodeid(mg, our_nodeid);
 
 	memset(buf, 0, MAXLINE);
 
@@ -1963,9 +1966,15 @@
 
 	if (error) {
 		log_group(mg, "leaving due to mount error: %s", mg->error_msg);
-		group_leave(gh, mg->name);
-	} else
+		if (memb->finished)
+			group_leave(gh, mg->name);
+		else {
+			log_group(mg, "delay leave until after join");
+			mg->group_leave_on_finish = 1;
+		}
+	} else {
 		mg->mount_client_notified = 1;
+	}
 }
 
 void ping_kernel_mount(char *table)
@@ -2192,6 +2201,13 @@
 	list_for_each_entry(memb, &mg->members, list)
 		memb->finished = 1;
 
+	if (mg->group_leave_on_finish) {
+		log_group(mg, "leaving group after delay for join to finish");
+		group_leave(gh, mg->name);
+		mg->group_leave_on_finish = 0;
+		return 0;
+	}
+
 	if (mg->needs_recovery) {
 		log_group(mg, "finish: leave locks blocked for needs_recovery");
 		leave_blocked = 1;