Global heartbeat - drop#1

Signed-off-by Sunil Mushran <[email protected]>

---
 fs/ocfs2/cluster/ocfs2_nodemanager.h |    2 +
 fs/ocfs2/dlm/dlmcommon.h             |   12 ++-
 fs/ocfs2/dlm/dlmdomain.c             |  222 +++++++++++++++++++++++++++++++++-
 3 files changed, 234 insertions(+), 2 deletions(-)

fs/ocfs2/cluster/ocfs2_nodemanager.h +2 -0

index 5b9854b..1829c01 100644
@@ -36,4 +36,6 @@
 /* host name, group name, cluster name all 64 bytes */
 #define O2NM_MAX_NAME_LEN        64    // __NEW_UTS_LEN
 
+#define O2NM_MAX_HBREGIONS	16
+
 #endif /* _OCFS2_NODEMANAGER_H */

fs/ocfs2/dlm/dlmcommon.h +11 -1

index a13292a..2c05138 100644
@@ -445,7 +445,8 @@ enum {
 	DLM_LOCK_REQUEST_MSG,	 /* 515 */
 	DLM_RECO_DATA_DONE_MSG,	 /* 516 */
 	DLM_BEGIN_RECO_MSG,	 /* 517 */
-	DLM_FINALIZE_RECO_MSG	 /* 518 */
+	DLM_FINALIZE_RECO_MSG,	 /* 518 */
+	DLM_QUERY_HBREGION,	 /* 519 */
 };
 
 struct dlm_reco_node_data

@@ -727,6 +728,15 @@ struct dlm_cancel_join
 	u8 domain[O2NM_MAX_NAME_LEN];
 };
 
+struct dlm_query_hbregion {
+	u8 qhb_node;
+	u8 qhb_numregions;
+	u8 qhb_namelen;
+	u8 pad1;
+	u8 qhb_domain[O2NM_MAX_NAME_LEN];
+	u8 qhb_hbregions[O2HB_MAX_REGION_NAME_LEN * O2NM_MAX_HBREGIONS];
+};
+
 struct dlm_exit_domain
 {
 	u8 node_idx;

fs/ocfs2/dlm/dlmdomain.c +81 -1

index 2408b9f..3521a00 100644
@@ -128,10 +128,13 @@ static DECLARE_WAIT_QUEUE_HEAD(dlm_domain_events);
  * will have a negotiated version with the same major number and a minor
  * number equal or smaller.  The dlm_ctxt->dlm_locking_proto field should
  * be used to determine what a running domain is actually using.
+ *
+ * New in version 1.1:
+ *	- Message DLM_QUERY_HBREGION added to support global heartbeat
  */
 static const struct dlm_protocol_version dlm_protocol = {
 	.pv_major = 1,
-	.pv_minor = 0,
+	.pv_minor = 1,
 };
 
 #define DLM_DOMAIN_BACKOFF_MS 200

@@ -142,6 +145,8 @@ static int dlm_assert_joined_handler(struct o2net_msg *msg, u32 len, void *data,
 				     void **ret_data);
 static int dlm_cancel_join_handler(struct o2net_msg *msg, u32 len, void *data,
 				   void **ret_data);
+static int dlm_query_hbregion_handler(struct o2net_msg *msg, u32 len,
+				      void *data, void **ret_data);
 static int dlm_exit_domain_handler(struct o2net_msg *msg, u32 len, void *data,
 				   void **ret_data);
 static int dlm_protocol_compare(struct dlm_protocol_version *existing,

@@ -918,6 +923,205 @@ static int dlm_assert_joined_handler(struct o2net_msg *msg, u32 len, void *data,
 	return 0;
 }
 
+static int dlm_match_hbregions(struct dlm_ctxt *dlm,
+			       struct dlm_query_hbregion *qhb)
+{
+	char *local = NULL, *remote = qhb->qhb_hbregions;
+	char *l, *r;
+	int localnr, i, j, foundit;
+	int status = 0;
+
+	if (!o2hb_global_heartbeat_active()) {
+		if (qhb->qhb_numregions) {
+			mlog(ML_ERROR, "Domain %s: Joining node %d has global "
+			     "heartbeat enabled but local node %d does not\n",
+			     qhb->qhb_domain, qhb->qhb_node, dlm->node_num);
+			status = -EINVAL;
+		}
+		goto bail;
+	}
+
+	if (o2hb_global_heartbeat_active() && !qhb->qhb_numregions) {
+		mlog(ML_ERROR, "Domain %s: Local node %d has global "
+		     "heartbeat enabled but joining node %d does not\n",
+		     qhb->qhb_domain, dlm->node_num, qhb->qhb_node);
+		status = -EINVAL;
+		goto bail;
+	}
+
+	r = remote;
+	for (i = 0; i < qhb->qhb_numregions; ++i) {
+		mlog(ML_NOTICE, "Region %.*s\n", O2HB_MAX_REGION_NAME_LEN, r);
+		r += O2HB_MAX_REGION_NAME_LEN;
+	}
+
+	local = kmalloc(sizeof(qhb->qhb_hbregions), GFP_KERNEL);
+	if (!local) {
+		status = -ENOMEM;
+		goto bail;
+	}
+
+	localnr = o2hb_get_all_regions(local, O2NM_MAX_HBREGIONS);
+
+	/* compare local regions with remote */
+	l = local;
+	for (i = 0; i < localnr; ++i) {
+		foundit = 0;
+		r = remote;
+		for (j = 0; j <= qhb->qhb_numregions; ++j) {
+			if (!memcmp(l, r, O2HB_MAX_REGION_NAME_LEN)) {
+				foundit = 1;
+				break;
+			}
+			r += O2HB_MAX_REGION_NAME_LEN;
+		}
+		if (!foundit) {
+			status = -EINVAL;
+			mlog(ML_ERROR, "Domain %s: Region '%.*s' registered "
+			     "in local node %d but not in joining node %d\n",
+			     qhb->qhb_domain, O2HB_MAX_REGION_NAME_LEN, l,
+			     dlm->node_num, qhb->qhb_node);
+			goto bail;
+		}
+		l += O2HB_MAX_REGION_NAME_LEN;
+	}
+
+	/* compare remote with local regions */
+	r = remote;
+	for (i = 0; i < qhb->qhb_numregions; ++i) {
+		foundit = 0;
+		l = local;
+		for (j = 0; j < localnr; ++j) {
+			if (!memcmp(r, l, O2HB_MAX_REGION_NAME_LEN)) {
+				foundit = 1;
+				break;
+			}
+			l += O2HB_MAX_REGION_NAME_LEN;
+		}

fs/ocfs2/dlm/dlmdomain.c +35 -0

+
+bail:
+	kfree(local);
+
+	return status;
+}
+
+static int dlm_send_hbregions(struct dlm_ctxt *dlm, unsigned long *node_map)
+{
+	struct dlm_query_hbregion *qhb = NULL;
+	int status, ret = 0, i;
+	char *p;
+
+	if (find_next_bit(node_map, O2NM_MAX_NODES, 0) >= O2NM_MAX_NODES)
+		goto bail;
+
+	qhb = kmalloc(sizeof(struct dlm_query_hbregion), GFP_KERNEL);
+	if (!qhb) {
+		ret = -ENOMEM;
+		mlog_errno(ret);
+		goto bail;
+	}
+
+	memset(qhb, 0, sizeof(struct dlm_query_hbregion));
+
+	qhb->qhb_node = dlm->node_num;
+	qhb->qhb_namelen = strlen(dlm->name);
+	memcpy(qhb->qhb_domain, dlm->name, qhb->qhb_namelen);
+	/* if local hb, the numregions will be zero */
+	if (o2hb_global_heartbeat_active())
+		qhb->qhb_numregions = o2hb_get_all_regions(qhb->qhb_hbregions,
+							   O2NM_MAX_HBREGIONS);
+
+	p = qhb->qhb_hbregions;
+	for (i = 0; i < qhb->qhb_numregions; ++i, p += O2HB_MAX_REGION_NAME_LEN)

fs/ocfs2/dlm/dlmdomain.c +3 -0

+		ret = o2net_send_message(DLM_QUERY_HBREGION, DLM_MOD_KEY, qhb,
+					 sizeof(struct dlm_query_hbregion),
+					 i, &status);

Hi Sunil,

Why don't we merge the two new message to the existing DLM_QUERY_JOIN_MSG? Because it will become too large?

Quoted 82 lines

fs/ocfs2/dlm/dlmdomain.c +80 -0

+static int dlm_match_hbregions(struct dlm_ctxt *dlm,
+			       struct dlm_query_hbregion *qhb)
+{
+	char *local = NULL, *remote = qhb->qhb_hbregions;
+	char *l, *r;
+	int localnr, i, j, foundit;
+	int status = 0;
+
+	if (!o2hb_global_heartbeat_active()) {
+		if (qhb->qhb_numregions) {
+			mlog(ML_ERROR, "Domain %s: Joining node %d has global "
+			     "heartbeat enabled but local node %d does not\n",
+			     qhb->qhb_domain, qhb->qhb_node, dlm->node_num);
+			status = -EINVAL;
+		}
+		goto bail;
+	}
+
+	if (o2hb_global_heartbeat_active()&&  !qhb->qhb_numregions) {
+		mlog(ML_ERROR, "Domain %s: Local node %d has global "
+		     "heartbeat enabled but joining node %d does not\n",
+		     qhb->qhb_domain, dlm->node_num, qhb->qhb_node);
+		status = -EINVAL;
+		goto bail;
+	}
+
+	r = remote;
+	for (i = 0; i<  qhb->qhb_numregions; ++i) {
+		mlog(ML_NOTICE, "Region %.*s\n", O2HB_MAX_REGION_NAME_LEN, r);
+		r += O2HB_MAX_REGION_NAME_LEN;
+	}
+
+	local = kmalloc(sizeof(qhb->qhb_hbregions), GFP_KERNEL);
+	if (!local) {
+		status = -ENOMEM;
+		goto bail;
+	}
+
+	localnr = o2hb_get_all_regions(local, O2NM_MAX_HBREGIONS);
+
+	/* compare local regions with remote */
+	l = local;
+	for (i = 0; i<  localnr; ++i) {
+		foundit = 0;
+		r = remote;
+		for (j = 0; j<= qhb->qhb_numregions; ++j) {
+			if (!memcmp(l, r, O2HB_MAX_REGION_NAME_LEN)) {
+				foundit = 1;
+				break;
+			}
+			r += O2HB_MAX_REGION_NAME_LEN;
+		}
+		if (!foundit) {
+			status = -EINVAL;
+			mlog(ML_ERROR, "Domain %s: Region '%.*s' registered "
+			     "in local node %d but not in joining node %d\n",
+			     qhb->qhb_domain, O2HB_MAX_REGION_NAME_LEN, l,
+			     dlm->node_num, qhb->qhb_node);
+			goto bail;
+		}
+		l += O2HB_MAX_REGION_NAME_LEN;
+	}
+
+	/* compare remote with local regions */
+	r = remote;
+	for (i = 0; i<  qhb->qhb_numregions; ++i) {
+		foundit = 0;
+		l = local;
+		for (j = 0; j<  localnr; ++j) {
+			if (!memcmp(r, l, O2HB_MAX_REGION_NAME_LEN)) {
+				foundit = 1;
+				break;
+			}
+			l += O2HB_MAX_REGION_NAME_LEN;
+		}
+		if (!foundit) {
+			status = -EINVAL;
+			mlog(ML_ERROR, "Domain %s: Region '%.*s' registered "
+			     "in joining node %d but not in local node %d\n",
+			     qhb->qhb_domain, O2HB_MAX_REGION_NAME_LEN, r,

fs/ocfs2/dlm/dlmdomain.c static int dlm_assert_joined_handler(struct o2net_msg *msg, u32 len, void *data, · line 1010

Quoted 6 lines

fs/ocfs2/dlm/dlmdomain.c +5 -0

+			     qhb->qhb_node, dlm->node_num);
+			goto bail;
+		}
+		r += O2HB_MAX_REGION_NAME_LEN;
+	}

Why need to compare again? just checking "qhb->qhb_numregions == localnr" is not fine?

Quoted 40 lines

fs/ocfs2/dlm/dlmdomain.c +40 -0

+
+bail:
+	kfree(local);
+
+	return status;
+}
+
+static int dlm_send_hbregions(struct dlm_ctxt *dlm, unsigned long *node_map)
+{
+	struct dlm_query_hbregion *qhb = NULL;
+	int status, ret = 0, i;
+	char *p;
+
+	if (find_next_bit(node_map, O2NM_MAX_NODES, 0)>= O2NM_MAX_NODES)
+		goto bail;
+
+	qhb = kmalloc(sizeof(struct dlm_query_hbregion), GFP_KERNEL);
+	if (!qhb) {
+		ret = -ENOMEM;
+		mlog_errno(ret);
+		goto bail;
+	}
+
+	memset(qhb, 0, sizeof(struct dlm_query_hbregion));
+
+	qhb->qhb_node = dlm->node_num;
+	qhb->qhb_namelen = strlen(dlm->name);
+	memcpy(qhb->qhb_domain, dlm->name, qhb->qhb_namelen);
+	/* if local hb, the numregions will be zero */
+	if (o2hb_global_heartbeat_active())
+		qhb->qhb_numregions = o2hb_get_all_regions(qhb->qhb_hbregions,
+							   O2NM_MAX_HBREGIONS);
+
+	p = qhb->qhb_hbregions;
+	for (i = 0; i<  qhb->qhb_numregions; ++i, p += O2HB_MAX_REGION_NAME_LEN)
+		mlog(ML_NOTICE, "Region %.*s\n", O2HB_MAX_REGION_NAME_LEN, p);
+
+	i = -1;
+	while ((i = find_next_bit(node_map, O2NM_MAX_NODES,
+				  i + 1))<  O2NM_MAX_NODES) {

fs/ocfs2/dlm/dlmdomain.c static int dlm_assert_joined_handler(struct o2net_msg *msg, u32 len, void *data, · line 1055

Quoted 6 lines

fs/ocfs2/dlm/dlmdomain.c +5 -0

+		if (i == dlm->node_num)
+			continue;
+
+		mlog(ML_NOTICE, "Sending hbregion to node %d\n", i);
+

Is this(also the aboves and the belows) NOTICE log needed? Guessing you were using them for debug purpose :-P

regards,
wengang.

Quoted 4 lines

fs/ocfs2/dlm/dlmdomain.c +3 -0

+		ret = o2net_send_message(DLM_QUERY_HBREGION, DLM_MOD_KEY, qhb,
+					 sizeof(struct dlm_query_hbregion),
+					 i,&status);

Thanks for your review.

Quoted 2 lines · Wengang Wang

Hi Sunil,

Quoted 2 lines

Why don't we merge the two new message to the existing DLM_QUERY_JOIN_MSG? Because it will become too large?

The two structures are very different. I see no advantage in us merging the two. Note that these messages are sent only during join. So saving a message round trip is not worth it.... especially if it means having one larger message.

Quoted 46 lines · Sunil Mushran

Quoted 44 lines

fs/ocfs2/dlm/dlmdomain.c +43 -0

+static int dlm_match_hbregions(struct dlm_ctxt *dlm,
+			       struct dlm_query_hbregion *qhb)
+{
+
+	/* compare local regions with remote */
+	l = local;
+	for (i = 0; i<  localnr; ++i) {
+		foundit = 0;
+		r = remote;
+		for (j = 0; j<= qhb->qhb_numregions; ++j) {
+			if (!memcmp(l, r, O2HB_MAX_REGION_NAME_LEN)) {
+				foundit = 1;
+				break;
+			}
+			r += O2HB_MAX_REGION_NAME_LEN;
+		}
+		if (!foundit) {
+			status = -EINVAL;
+			mlog(ML_ERROR, "Domain %s: Region '%.*s' registered "
+			     "in local node %d but not in joining node %d\n",
+			     qhb->qhb_domain, O2HB_MAX_REGION_NAME_LEN, l,
+			     dlm->node_num, qhb->qhb_node);
+			goto bail;
+		}
+		l += O2HB_MAX_REGION_NAME_LEN;
+	}
+
+	/* compare remote with local regions */
+	r = remote;
+	for (i = 0; i<  qhb->qhb_numregions; ++i) {
+		foundit = 0;
+		l = local;
+		for (j = 0; j<  localnr; ++j) {
+			if (!memcmp(r, l, O2HB_MAX_REGION_NAME_LEN)) {
+				foundit = 1;
+				break;
+			}
+			l += O2HB_MAX_REGION_NAME_LEN;
+		}
+		if (!foundit) {
+			status = -EINVAL;
+			mlog(ML_ERROR, "Domain %s: Region '%.*s' registered "
+			     "in joining node %d but not in local node %d\n",

Quoted 8 lines

fs/ocfs2/dlm/dlmdomain.c static int dlm_assert_joined_handler(struct o2net_msg *msg, u32 len, void *data, · line 1010

Quoted 6 lines

fs/ocfs2/dlm/dlmdomain.c +6 -0

+			     qhb->qhb_domain, O2HB_MAX_REGION_NAME_LEN, r,
+			     qhb->qhb_node, dlm->node_num);
+			goto bail;
+		}
+		r += O2HB_MAX_REGION_NAME_LEN;
+	}

Why need to compare again? just checking "qhb->qhb_numregions == localnr" is not fine?

Say the remote has hbregions A and B and local has C and A. We want all the regions to match. And they can be presented in any order. I am just going with brute force for now. Remember we allow a max of 16 regions.

Yes, I can see what you are doing there.
But you are comparing twice. I was emphasising "again" :) The bhregions is like a collection. we say collection A is equal to collection B, it can mean the number is equal, and all elements in collection A are all in collection B. So no need to compare each region again.

Quoted 6 lines · Wengang Wang

Yes, I can see what you are doing there.
But you are comparing twice. I was emphasising "again" :) The bhregions is like a collection. we say collection A is equal to collection B, it can mean the number is equal, and all elements in collection A are all in collection B. So no need to compare each region again.

Sorry, my example was not correct. What if node sends regions A, B, A while the receiving node expects A, B, C. It should not happen. But considering we rely on the hb regions being consistent, we can never be too careful.

Agreed, I can come up with another scheme. But here the max is small. And in such cases, brute force typically works best.

static int dlm_cancel_join_handler(struct o2net_msg *msg, u32 len, void *data,

fs/ocfs2/dlm/dlmdomain.c +0 -0

 				   void **ret_data)
 {

@@ -1233,6 +1437,15 @@ static int dlm_try_to_join_domain(struct dlm_ctxt *dlm)
 	set_bit(dlm->node_num, dlm->domain_map);
 	spin_unlock(&dlm->spinlock);

fs/ocfs2/dlm/dlmdomain.c +0 -0

 	dlm_send_join_asserts(dlm, ctxt->yes_resp_map);
 

Quoted 7 lines

static int dlm_cancel_join_handler(struct o2net_msg *msg, u32 len, void *data,
void **ret_data)
{

fs/ocfs2/dlm/dlmdomain.c +0 -0

@@ -1233,6 +1437,15 @@ static int dlm_try_to_join_domain(struct dlm_ctxt *dlm)
  	set_bit(dlm->node_num, dlm->domain_map);
  	spin_unlock(&dlm->spinlock);
 

fs/ocfs2/dlm/dlmdomain.c static int dlm_try_to_join_domain(struct dlm_ctxt *dlm) · line 1449

Quoted 10 lines

fs/ocfs2/dlm/dlmdomain.c +9 -0

+	/* Support for global heartbeat was added in 1.1 */
+	if (dlm_protocol.pv_major>= 1&&  dlm_protocol.pv_minor>  0) {
+		status = dlm_send_hbregions(dlm, ctxt->yes_resp_map);
+		if (status) {
+			mlog_errno(status);
+			goto bail;
+		}
+	}
+
 

How does the version number grows? Is it possible from 1.1 to 2.0? If that's possible, should it be

Signed-off-by Sunil Mushran <[email protected]>

---
 fs/ocfs2/dlm/dlmcommon.h |   17 ++++
 fs/ocfs2/dlm/dlmdomain.c |  188 +++++++++++++++++++++++++++++++++++++++++++++-
 2 files changed, 204 insertions(+), 1 deletions(-)

fs/ocfs2/dlm/dlmdomain.c +8 -0

+	for (i = 0; i < O2NM_MAX_NODES && !status; ++i) {
+		local = o2nm_get_node_by_num(i);
+		remote = NULL;
+		for (j = 0; j < qn->qn_numnodes; ++j) {
+			if (qn->qn_nodes[j].ni_nodenum == i) {
+				remote = &(qn->qn_nodes[j]);
+				break;
+			}

fs/ocfs2/dlm/dlmdomain.c +11 -0

+		     (remote->ni_ipv4_port != local->nd_ipv4_port) ||
+		     (remote->ni_ipv4_address != local->nd_ipv4_address)))
+			status = -EINVAL;
+
 
+static int dlm_send_nodeinfo(struct dlm_ctxt *dlm, unsigned long *node_map)
+{
+	struct dlm_query_nodeinfo *qn = NULL;
+	struct o2nm_node *node;
+	int ret = 0, status, count, i;
+
+	if (find_next_bit(node_map, O2NM_MAX_NODES, 0) >= O2NM_MAX_NODES)

fs/ocfs2/dlm/dlmdomain.c +10 -0

+	for (i = 0, count = 0; i < O2NM_MAX_NODES; ++i) {
+		node = o2nm_get_node_by_num(i);
+		if (!node)
+			continue;
+		qn->qn_nodes[count].ni_nodenum = node->nd_num;
+		qn->qn_nodes[count].ni_ipv4_port = node->nd_ipv4_port;
+
 
+static int dlm_query_nodeinfo_handler(struct o2net_msg *msg, u32 len,
+				      void *data, void **ret_data)
+{

fs/ocfs2/dlm/dlmdomain.c +9 -0

+
+	spin_lock(&dlm_domain_lock);
+	dlm = __dlm_lookup_domain_full(qn->qn_domain, qn->qn_namelen);
+	if (!dlm) {
+		mlog(ML_ERROR, "Node %d queried nodes on domain %s before "
+		     "join domain\n", qn->qn_nodenum, qn->qn_domain);
+		goto bail;
+	}
+

ni_nodenum already checked in the for loop.

Any reason we are not using kzalloc() here and in dlm_send_hbregions()?

This is the first drop of the global heartbeat patches for ocfs2/kernel.

The first few patches add support for heartbeat mode in sysfs, the new incompat clusterinfo flag and the new mount option heartbeat=global.

0001-ocfs2-cluster-Add-heartbeat-mode-configfs-parameter.patch 0002-ocfs2-Add-an-incompat-feature-flag-OCFS2_FEATURE_INC.patch 0003-ocfs2-Add-support-for-heartbeat-global-mount-option.patch 0004-ocfs2-dlm-Expose-dlm_protocol-in-dlm_state.patch

The next few patches enhance the join domain protocol to get the list of configured nodes and heartbeating regions to ensure that all nodes in the cluster have the same view of the cluster.

0005-ocfs2-cluster-Get-all-heartbeat-regions.patch 0006-ocfs2-dlm-Add-message-DLM_QUERY_HBREGION.patch 0007-ocfs2-Print-message-if-user-mounts-without-starting-.patch 0008-ocfs2-dlm-Add-message-DLM_QUERY_NODEINFO.patch

The one known missing bit concerns quorum calculation. I am still working on it.

http://oss.oracle.com/osswiki/OCFS2/DesignDocs/NewGlobalHeartbeat

Thanks
Sunil

_______________________________________________
Ocfs2-devel mailing list
Ocfs2-devel at oss.oracle.com
http://oss.oracle.com/mailman/listinfo/ocfs2-devel

Hi Sunil,

The global heartbeat also introduce a difference comparing with local heart. With the ghb, what if the non-heartbeat ocfs2 volume(s) fail(s)? Say some lower layer(raid/disk-driver) become unhappy to work anymore. In case, no failover since no self-fence I think. So it could cause the domain in question unavailable(finally), isn't it? With the original lhb scheme, no such problem. Is there a solution?

regards,
wengang.

Quoted 32 lines · Sunil Mushran

This is the first drop of the global heartbeat patches for ocfs2/kernel.

The first few patches add support for heartbeat mode in sysfs, the new incompat clusterinfo flag and the new mount option heartbeat=global.

0001-ocfs2-cluster-Add-heartbeat-mode-configfs-parameter.patch 0002-ocfs2-Add-an-incompat-feature-flag-OCFS2_FEATURE_INC.patch 0003-ocfs2-Add-support-for-heartbeat-global-mount-option.patch 0004-ocfs2-dlm-Expose-dlm_protocol-in-dlm_state.patch

The next few patches enhance the join domain protocol to get the list of configured nodes and heartbeating regions to ensure that all nodes in the cluster have the same view of the cluster.

0005-ocfs2-cluster-Get-all-heartbeat-regions.patch 0006-ocfs2-dlm-Add-message-DLM_QUERY_HBREGION.patch 0007-ocfs2-Print-message-if-user-mounts-without-starting-.patch 0008-ocfs2-dlm-Add-message-DLM_QUERY_NODEINFO.patch

The one known missing bit concerns quorum calculation. I am still working on it.

http://oss.oracle.com/osswiki/OCFS2/DesignDocs/NewGlobalHeartbeat

Thanks
Sunil

_______________________________________________
Ocfs2-devel mailing list
Ocfs2-devel at oss.oracle.com
http://oss.oracle.com/mailman/listinfo/ocfs2-devel

If the io stack on a non-hb device fails, the apps will start getting EIOs. Possibly leading to application death. But the system will stay up.

The dlm domain should remain available.

One case I can see happening is if the journal op fail (say commit triggered by downconvert). In that case, the fs will fence the box.

Do you have a specific example?

Quoted 9 lines

Hi Sunil,

The global heartbeat also introduce a difference comparing with local heart. With the ghb, what if the non-heartbeat ocfs2 volume(s) fail(s)? Say some lower layer(raid/disk-driver) become unhappy to work anymore. In case, no failover since no self-fence I think. So it could cause the domain in question unavailable(finally), isn't it? With the original lhb scheme, no such problem. Is there a solution?

Quoted 6 lines

One case I can see happening is if the journal op fail (say commit triggered by downconvert). In that case, the fs will fence the box.

Do you have a specific example?

I have no good example, but just thinking about this case:

A two nodes HA cluster, node A is active and node B is backup.
If the non-hb device fails on node A, since no failover will happen in this case, the app will keep getting the EIOs on the active node. User maybe complain with that? With lhb, node A will be restarted and failover will happen so that user app won't keep getting EIOs.

Or all nodes are active nodes, but they all get EIOs since no fence in this case.

It is just a question, not a big problem for the helpful ghb :)

Quoted 13 lines · Wengang Wang

Quoted 5 lines

One case I can see happening is if the journal op fail (say commit triggered by downconvert). In that case, the fs will fence the box.

Do you have a specific example?

I have no good example, but just thinking about this case:

A two nodes HA cluster, node A is active and node B is backup.
If the non-hb device fails on node A, since no failover will happen in this case, the app will keep getting the EIOs on the active node. User maybe complain with that? With lhb, node A will be restarted and failover will happen so that user app won't keep getting EIOs.

Quoted 4 lines

Or all nodes are active nodes, but they all get EIOs since no fence in this case.

It is just a question, not a big problem for the helpful ghb :)

o2hb is meant only for ocfs2. If one wants app failover, then one can use clusterware or rgmanager or pacemaker. Or one can hand craft app failover on ocfs2 using dlmfs. I know users that are doing so.