Skip to contentSkip to conversation

ocfs2/cluster: Make fence method configurable

PATCH New Sunil Mushran 2 participants 2 messages Nov 17, 2009 OCFS2

lore
Share & tools
lore
#0 lore raw

PATCH 1/1 ocfs2/cluster: Make fence method configurable

By default, o2cb fences the box by calling emergency_restart(). While this scheme works well in production, it comes in the way during testing as it does not let the tester take stack/core dumps for analysis.

This patch allows user to dynamically change the fence method to panic() by:

# echo "panic" > /sys/kernel/config/cluster/<clustername>/fence_method
---
 fs/ocfs2/cluster/nodemanager.c |   50 ++++++++++++++++++++++++++++++++++++++++
 fs/ocfs2/cluster/nodemanager.h |    7 +++++
 fs/ocfs2/cluster/quorum.c      |   14 +++++++++-
 3 files changed, 69 insertions(+), 2 deletions(-)
fs/ocfs2/cluster/nodemanager.c +50 -0
index 7ee6188..77cdc09 100644
@@ -35,6 +35,11 @@
  * cluster references throughout where nodes are looked up */
 struct o2nm_cluster *o2nm_single_cluster = NULL;
 
+char *o2nm_fence_method_desc[O2NM_FENCE_METHODS + 1] = {
+		"reset",	/* O2NM_FENCE_RESET */
+		"panic",	/* O2NM_FENCE_PANIC */
+		"unknown",
+};
 
 struct o2nm_node *o2nm_get_node_by_num(u8 node_num)
 {
@@ -579,6 +584,41 @@ static ssize_t o2nm_cluster_attr_reconnect_delay_ms_write(
 	return o2nm_cluster_attr_write(page, count,
 	                               &cluster->cl_reconnect_delay_ms);
 }
+
+static ssize_t o2nm_cluster_attr_fence_method_read(
+	struct o2nm_cluster *cluster, char *page)
+{
+	unsigned int i = O2NM_FENCE_METHODS;
+
+	if (cluster && cluster->cl_fence_method < O2NM_FENCE_METHODS)
+		i = cluster->cl_fence_method;
+
+	return sprintf(page, "%s\n", o2nm_fence_method_desc[i]);
+}
+
+static ssize_t o2nm_cluster_attr_fence_method_write(
+	struct o2nm_cluster *cluster, const char *page, size_t count)
+{
+	unsigned int i;
+
+	if (page[count - 1] != '\n')
+		goto bail;
+
+	for (i = 0; i < O2NM_FENCE_METHODS; ++i) {
+		if (count != strlen(o2nm_fence_method_desc[i]) + 1)
+		    continue;
+		if (!strncasecmp(page, o2nm_fence_method_desc[i], count - 1)) {
+			printk(KERN_INFO "ocfs2: Changing fence method to %s\n",
+			       o2nm_fence_method_desc[i]);
+			cluster->cl_fence_method = i;
+			return count;
+		}
+	}
+
+bail:
+	return -EINVAL;
+}
+
 static struct o2nm_cluster_attribute o2nm_cluster_attr_idle_timeout_ms = {
 	.attr	= { .ca_owner = THIS_MODULE,
 		    .ca_name = "idle_timeout_ms",
@@ -603,10 +643,19 @@ static struct o2nm_cluster_attribute o2nm_cluster_attr_reconnect_delay_ms = {
 	.store	= o2nm_cluster_attr_reconnect_delay_ms_write,
 };
 
+static struct o2nm_cluster_attribute o2nm_cluster_attr_fence_method = {
+	.attr	= { .ca_owner = THIS_MODULE,
+		    .ca_name = "fence_method",
+		    .ca_mode = S_IRUGO | S_IWUSR },
+	.show	= o2nm_cluster_attr_fence_method_read,
+	.store	= o2nm_cluster_attr_fence_method_write,
+};
+
 static struct configfs_attribute *o2nm_cluster_attrs[] = {
 	&o2nm_cluster_attr_idle_timeout_ms.attr,
 	&o2nm_cluster_attr_keepalive_delay_ms.attr,
 	&o2nm_cluster_attr_reconnect_delay_ms.attr,
+	&o2nm_cluster_attr_fence_method.attr,
 	NULL,
 };
 static ssize_t o2nm_cluster_show(struct config_item *item,
@@ -778,6 +827,7 @@ static struct config_group *o2nm_cluster_group_make_group(struct config_group *g
 	cluster->cl_reconnect_delay_ms = O2NET_RECONNECT_DELAY_MS_DEFAULT;
 	cluster->cl_idle_timeout_ms    = O2NET_IDLE_TIMEOUT_MS_DEFAULT;
 	cluster->cl_keepalive_delay_ms = O2NET_KEEPALIVE_DELAY_MS_DEFAULT;
+	cluster->cl_fence_method       = O2NM_FENCE_RESET;
 
 	ret = &cluster->cl_group;
 	o2nm_single_cluster = cluster;
fs/ocfs2/cluster/nodemanager.h +7 -0
index c992ea0..09ea2d3 100644
@@ -33,6 +33,12 @@
 #include <linux/configfs.h>
 #include <linux/rbtree.h>
 
+enum o2nm_fence_method {
+	O2NM_FENCE_RESET	= 0,
+	O2NM_FENCE_PANIC,
+	O2NM_FENCE_METHODS,	/* Number of fence methods */
+};
+
 struct o2nm_node {
 	spinlock_t		nd_lock;
 	struct config_item	nd_item;
@@ -58,6 +64,7 @@ struct o2nm_cluster {
 	unsigned int		cl_idle_timeout_ms;
 	unsigned int		cl_keepalive_delay_ms;
 	unsigned int		cl_reconnect_delay_ms;
+	enum o2nm_fence_method	cl_fence_method;
 
 	/* this bitmap is part of a hack for disk bitmap.. will go eventually. - zab */
 	unsigned long	cl_nodes_bitmap[BITS_TO_LONGS(O2NM_MAX_NODES)];
fs/ocfs2/cluster/quorum.c +12 -2
index bbacf7d..cc6ed4e 100644
@@ -74,8 +74,18 @@ static void o2quo_fence_self(void)
 	 * threads can still schedule, etc, etc */
 	o2hb_stop_all_regions();
 
-	printk("ocfs2 is very sorry to be fencing this system by restarting\n");
-	emergency_restart();
+	switch (o2nm_single_cluster->cl_fence_method) {
+	case O2NM_FENCE_PANIC:
+		panic("*** ocfs2 is very sorry to be fencing this system by "
+		      "panicing ***\n");
+		break;
+	case O2NM_FENCE_RESET:
+	default:
+		printk("*** ocfs2 is very sorry to be fencing this system by "
+		       "restarting ***\n");
+		emergency_restart();
+		break;
+	};
 }
 
 /* Indicate that a timeout occured on a hearbeat region write. The
Signature · 1 line
1.5.6.5
#1 lore raw

PATCH 1/1 ocfs2/cluster: Make fence method configurable

Quoted 19 lines · Sunil Mushran

By default, o2cb fences the box by calling emergency_restart(). While this scheme works well in production, it comes in the way during testing as it does not let the tester take stack/core dumps for analysis.

This patch allows user to dynamically change the fence method to panic() by:

# echo "panic" > /sys/kernel/config/cluster/<clustername>/fence_method
---
 fs/ocfs2/cluster/nodemanager.c |   50 ++++++++++++++++++++++++++++++++++++++++
 fs/ocfs2/cluster/nodemanager.h |    7 +++++
 fs/ocfs2/cluster/quorum.c      |   14 +++++++++-
 3 files changed, 69 insertions(+), 2 deletions(-)
fs/ocfs2/cluster/nodemanager.c +0 -0
index 7ee6188..77cdc09 100644

fs/ocfs2/cluster/nodemanager.c line 42

Quoted 8 lines
fs/ocfs2/cluster/nodemanager.c +5 -0
@@ -35,6 +35,11 @@
  * cluster references throughout where nodes are looked up */
 struct o2nm_cluster *o2nm_single_cluster = NULL;
 
+char *o2nm_fence_method_desc[O2NM_FENCE_METHODS + 1] = {
+		"reset",	/* O2NM_FENCE_RESET */
+		"panic",	/* O2NM_FENCE_PANIC */
+		"unknown",
+};

Why do you have "unknown"? There's no point to it; we
default to "reset" and you never allow anything about O2NM_FENCE_METHODS to be set.

fs/ocfs2/cluster/nodemanager.c static ssize_t o2nm_cluster_attr_reconnect_delay_ms_write( · line 594

Quoted 7 lines
fs/ocfs2/cluster/nodemanager.c +7 -0
+static ssize_t o2nm_cluster_attr_fence_method_read(
+	struct o2nm_cluster *cluster, char *page)
+{
+	unsigned int i = O2NM_FENCE_METHODS;
+
+	if (cluster && cluster->cl_fence_method < O2NM_FENCE_METHODS)
+		i = cluster->cl_fence_method;

cl_fence_method should always be below O2NM_FENCE_METHODS. You
only need to check if(cluster). If cl_fence_method is off, we'll crash, which is fine.

fs/ocfs2/cluster/nodemanager.c

Quoted 7 lines
fs/ocfs2/cluster/nodemanager.c +7 -0
+	switch (o2nm_single_cluster->cl_fence_method) {
+	case O2NM_FENCE_PANIC:
+		panic("*** ocfs2 is very sorry to be fencing this system by "
+		      "panicing ***\n");
+		break;
+	case O2NM_FENCE_RESET:
+	default:

Why not:

+ default:
+ WARN_ON(o2nm_single_cluster->cl_fence_method >=
+ O2NM_FENCE_METHODS);
+ case O2NM_FENCE_RESET:

so that netconsole will catch a wacky cl_fence_method.

fs/ocfs2/cluster/nodemanager.c

Quoted 2 lines
fs/ocfs2/cluster/nodemanager.c +2 -0
+		printk("*** ocfs2 is very sorry to be fencing this system by "
+		       "restarting ***\n");

And here, print KERN_ERR so noone filters it out with low klogd
levels.

Joel

Signature · 9 lines
"For every complex problem there exists a solution that is brief,
     concise, and totally wrong."
                                        -Unknown

Joel Becker
Principal Software Developer
Oracle
E-mail: joel.becker at oracle.com
Phone: (650) 506-8127