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

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

[teigland]

CVSROOT:	/cvs/cluster
Module name:	cluster
Changes by:	teigland at sourceware.org	2006-07-14 18:56:10

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

Log message:
	keep a 1MB circular buffer of debug messages, they can be dumped out
	by running 'group_tool dump gfs'

Patches:
http://sourceware.org/cgi-bin/cvsweb.cgi/cluster/group/gfs_controld/lock_dlm.h.diff?cvsroot=cluster&r1=1.4&r2=1.5
http://sourceware.org/cgi-bin/cvsweb.cgi/cluster/group/gfs_controld/main.c.diff?cvsroot=cluster&r1=1.4&r2=1.5

--- cluster/group/gfs_controld/lock_dlm.h	2006/07/13 21:28:45	1.4
+++ cluster/group/gfs_controld/lock_dlm.h	2006/07/14 18:56:10	1.5
@@ -43,6 +43,7 @@
 #define MAX_CLIENTS		8
 #define MAX_MSGLEN		2048
 #define MAX_OPTIONS_LEN		1024
+#define DUMP_SIZE		(1024 * 1024)
 
 #define LOCK_DLM_GROUP_LEVEL	2
 #define LOCK_DLM_GROUP_NAME	"gfs"
@@ -67,11 +68,17 @@
 extern char *prog_name;
 extern int daemon_debug_opt;
 extern char daemon_debug_buf[256];
+extern char dump_buf[DUMP_SIZE];
+extern int dump_point;
+extern int dump_wrap;
+
+extern void daemon_dump_save(void);
 
 #define log_debug(fmt, args...) \
 do { \
 	snprintf(daemon_debug_buf, 255, "%ld " fmt "\n", time(NULL), ##args); \
 	if (daemon_debug_opt) fprintf(stderr, "%s", daemon_debug_buf); \
+	daemon_dump_save(); \
 } while (0)
 
 #define log_group(g, fmt, args...) \
@@ -79,6 +86,7 @@
 	snprintf(daemon_debug_buf, 255, "%ld %s " fmt "\n", time(NULL), \
 		 (g)->name, ##args); \
 	if (daemon_debug_opt) fprintf(stderr, "%s", daemon_debug_buf); \
+	daemon_dump_save(); \
 } while (0)
 
 #define log_error(fmt, args...) \
--- cluster/group/gfs_controld/main.c	2006/07/13 21:28:45	1.4
+++ cluster/group/gfs_controld/main.c	2006/07/14 18:56:10	1.5
@@ -125,6 +125,24 @@
 	return write(client[ci].fd, buf, len);
 }
 
+static int do_dump(int ci)
+{
+	int rv, len;
+
+	if (dump_wrap) {
+		len = DUMP_SIZE - dump_point;
+		rv = write(client[ci].fd, dump_buf + dump_point, len);
+		if (rv != len)
+			log_debug("write error %d errno %d", rv, errno);
+	}
+	len = dump_point;
+
+	rv = write(client[ci].fd, dump_buf, len);
+	if (rv != len)
+		log_debug("write error %d errno %d", rv, errno);
+	return 0;
+}
+
 static int process_client(int ci)
 {
 	char *cmd, *dir, *type, *proto, *table, *extra;
@@ -163,7 +181,10 @@
 		rv = do_unmount(ci, dir);
 	else if (!strcmp(cmd, "remount"))
 		rv = do_remount(ci, dir, argv[3]);
-	else
+	else if (!strcmp(cmd, "dump")) {
+		do_dump(ci);
+		return 0;
+	} else
 		rv = -EINVAL;
 
 	sprintf(out, "%d", rv);
@@ -492,7 +513,26 @@
 	return loop();
 }
 
+void daemon_dump_save(void)
+{
+	int len, i;
+
+	len = strlen(daemon_debug_buf);
+
+	for (i = 0; i < len; i++) {
+		dump_buf[dump_point++] = daemon_debug_buf[i];
+
+		if (dump_point == DUMP_SIZE) {
+			dump_point = 0;
+			dump_wrap = 1;
+		}
+	}
+}
+
 char *prog_name;
 int daemon_debug_opt;
 char daemon_debug_buf[256];
+char dump_buf[DUMP_SIZE];
+int dump_point;
+int dump_wrap;
 

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

[teigland]

CVSROOT:	/cvs/cluster
Module name:	cluster
Branch: 	RHEL5
Changes by:	teigland at sourceware.org	2007-06-12 20:05:12

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

Log message:
	log an error message if we see mount.gfs killed before it's done

Patches:
http://sourceware.org/cgi-bin/cvsweb.cgi/cluster/group/gfs_controld/lock_dlm.h.diff?cvsroot=cluster&only_with_tag=RHEL5&r1=1.21.2.6&r2=1.21.2.7
http://sourceware.org/cgi-bin/cvsweb.cgi/cluster/group/gfs_controld/main.c.diff?cvsroot=cluster&only_with_tag=RHEL5&r1=1.18.2.11&r2=1.18.2.12

--- cluster/group/gfs_controld/lock_dlm.h	2006/12/20 19:14:41	1.21.2.6
+++ cluster/group/gfs_controld/lock_dlm.h	2007/06/12 20:05:12	1.21.2.7
@@ -289,6 +289,7 @@
 void got_mount_result(struct mountgroup *mg, int result, int ci, int another);
 
 int client_send(int ci, char *buf, int len);
+int get_sysfs(struct mountgroup *mg, char *field, char *buf, int len);
 
 int send_group_message(struct mountgroup *mg, int len, char *buf);
 void update_flow_control_status(void);
--- cluster/group/gfs_controld/main.c	2007/04/26 19:01:42	1.18.2.11
+++ cluster/group/gfs_controld/main.c	2007/06/12 20:05:12	1.18.2.12
@@ -39,6 +39,7 @@
 
 extern struct list_head mounts;
 extern struct list_head withdrawn_mounts;
+extern group_handle_t gh;
 int no_withdraw;
 int no_plock;
 uint32_t plock_rate_limit = DEFAULT_PLOCK_RATE_LIMIT;
@@ -171,9 +172,66 @@
 	}
 }
 
+/* I don't think we really want to try to do anything if mount.gfs is killed,
+   because I suspect there are various corner cases where we might not do the
+   right thing.  Even without the corner cases things still don't work out
+   too nicely.  Best to just tell people not to kill a mount or unmount
+   because doing so can leave things (kernel, group, mtab) in inconsistent
+   states that can't be straightened out properly without a reboot. */
+
+static void mount_client_dead(struct mountgroup *mg, int ci)
+{
+	char buf[MAXLINE];
+	int rv;
+
+	if (ci != mg->mount_client) {
+		log_error("mount client mismatch %d %d", ci, mg->mount_client);
+		return;
+	}
+
+	/* is checking sysfs really a reliable way of telling whether the
+	   kernel has been mounted or not?  might the kernel mount just not
+	   have reached the sysfs registration yet? */
+
+	memset(buf, 0, sizeof(buf));
+
+	rv = get_sysfs(mg, "id", buf, sizeof(buf));
+	if (!rv) {
+		log_error("mount_client_dead ci %d sysfs id %s", ci, buf);
+#if 0
+		/* finish the mount, although there will be no mtab entry
+		   which will confuse umount causing it to do the kernel
+		   umount but not call umount.gfs */
+		got_mount_result(mg, 0, ci, client[ci].another_mount);
+#endif
+		return;
+	}
+
+	log_error("mount_client_dead ci %d no sysfs entry for fs", ci);
+
+#if 0
+	mp = find_mountpoint_client(mg, ci);
+	if (mp) {
+		list_del(&mp->list);
+		free(mp);
+	}
+	group_leave(gh, mg->name);
+#endif
+}
+
 static void client_dead(int ci)
 {
+	struct mountgroup *mg;
+
 	log_debug("client %d fd %d dead", ci, client[ci].fd);
+
+	/* if the dead mount client is mount.gfs and we've not received
+	   a mount result, then try to put things into a clean state */
+	   
+	mg = client[ci].mg;
+	if (mg && mg->mount_client && mg->mount_client_fd)
+		mount_client_dead(mg, ci);
+
 	close(client[ci].fd);
 	client[ci].fd = -1;
 	pollfd[ci].fd = -1;

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

[teigland]

CVSROOT:	/cvs/cluster
Module name:	cluster
Changes by:	teigland at sourceware.org	2007-06-12 20:04:41

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

Log message:
	log an error message if we see mount.gfs killed before it's done

Patches:
http://sourceware.org/cgi-bin/cvsweb.cgi/cluster/group/gfs_controld/lock_dlm.h.diff?cvsroot=cluster&r1=1.27&r2=1.28
http://sourceware.org/cgi-bin/cvsweb.cgi/cluster/group/gfs_controld/main.c.diff?cvsroot=cluster&r1=1.29&r2=1.30

--- cluster/group/gfs_controld/lock_dlm.h	2006/12/20 19:13:13	1.27
+++ cluster/group/gfs_controld/lock_dlm.h	2007/06/12 20:04:41	1.28
@@ -289,6 +289,7 @@
 void got_mount_result(struct mountgroup *mg, int result, int ci, int another);
 
 int client_send(int ci, char *buf, int len);
+int get_sysfs(struct mountgroup *mg, char *field, char *buf, int len);
 
 int send_group_message(struct mountgroup *mg, int len, char *buf);
 void update_flow_control_status(void);
--- cluster/group/gfs_controld/main.c	2007/04/26 19:01:34	1.29
+++ cluster/group/gfs_controld/main.c	2007/06/12 20:04:41	1.30
@@ -39,6 +39,7 @@
 
 extern struct list_head mounts;
 extern struct list_head withdrawn_mounts;
+extern group_handle_t gh;
 int no_withdraw;
 int no_plock;
 uint32_t plock_rate_limit = DEFAULT_PLOCK_RATE_LIMIT;
@@ -171,9 +172,66 @@
 	}
 }
 
+/* I don't think we really want to try to do anything if mount.gfs is killed,
+   because I suspect there are various corner cases where we might not do the
+   right thing.  Even without the corner cases things still don't work out
+   too nicely.  Best to just tell people not to kill a mount or unmount
+   because doing so can leave things (kernel, group, mtab) in inconsistent
+   states that can't be straightened out properly without a reboot. */
+
+static void mount_client_dead(struct mountgroup *mg, int ci)
+{
+	char buf[MAXLINE];
+	int rv;
+
+	if (ci != mg->mount_client) {
+		log_error("mount client mismatch %d %d", ci, mg->mount_client);
+		return;
+	}
+
+	/* is checking sysfs really a reliable way of telling whether the
+	   kernel has been mounted or not?  might the kernel mount just not
+	   have reached the sysfs registration yet? */
+
+	memset(buf, 0, sizeof(buf));
+
+	rv = get_sysfs(mg, "id", buf, sizeof(buf));
+	if (!rv) {
+		log_error("mount_client_dead ci %d sysfs id %s", ci, buf);
+#if 0
+		/* finish the mount, although there will be no mtab entry
+		   which will confuse umount causing it to do the kernel
+		   umount but not call umount.gfs */
+		got_mount_result(mg, 0, ci, client[ci].another_mount);
+#endif
+		return;
+	}
+
+	log_error("mount_client_dead ci %d no sysfs entry for fs", ci);
+
+#if 0
+	mp = find_mountpoint_client(mg, ci);
+	if (mp) {
+		list_del(&mp->list);
+		free(mp);
+	}
+	group_leave(gh, mg->name);
+#endif
+}
+
 static void client_dead(int ci)
 {
+	struct mountgroup *mg;
+
 	log_debug("client %d fd %d dead", ci, client[ci].fd);
+
+	/* if the dead mount client is mount.gfs and we've not received
+	   a mount result, then try to put things into a clean state */
+	   
+	mg = client[ci].mg;
+	if (mg && mg->mount_client && mg->mount_client_fd)
+		mount_client_dead(mg, ci);
+
 	close(client[ci].fd);
 	client[ci].fd = -1;
 	pollfd[ci].fd = -1;