[pve-devel] corosync bug: cluster break after 1 node clean shutdown

Fabian Grünbichler f.gruenbichler at proxmox.com
Thu Sep 24 16:02:04 CEST 2020


On September 22, 2020 7:43 am, Alexandre DERUMIER wrote:
> I have done test with "kill -9 <pidofcorosync",  and I have around 20s hang on other nodes,
> but after that it's become available again.
> 
> 
> So, it's really something when corosync is in shutdown phase, and pmxcfs is running.
> 
> So, for now, as workaround, I have changed
> 
> /lib/systemd/system/pve-cluster.service
> 
> #Wants=corosync.service
> #Before=corosync.service
> Requires=corosync.service
> After=corosync.service
> 
> 
> Like this, at shutdown, pve-cluster is stopped before corosync, and if I restart corosync, pve-cluster is stopped first.

if you are still able to test, it would be great if you could give the 
following packages a spin (they only contain some extra debug prints 
on message processing/sending):

http://download.proxmox.com/temp/pmxcfs-dbg/

64eb9dbd2f60fe319abaeece89a84fd5de1f05a8c38cb9871058ec1f55025486ec15b7c0053976159fe5c2518615fd80084925abf4d3a5061ea7d6edef264c36  pve-cluster_6.1-8_amd64.deb
04b557c7f0dc1aa2846b534d6afab70c2b8d4720ac307364e015d885e2e997b6dcaa54cad673b22d626d27cb053e5723510fde15d078d5fe1f262fc5486e6cef  pve-cluster-dbgsym_6.1-8_amd64.deb

ideally, you could get the debug logs from all nodes, and the 
coredump/bt from the node where pmxcfs hangs. thanks!

diff --git a/data/src/dfsm.c b/data/src/dfsm.c
index 529c7f9..e0bd93f 100644
--- a/data/src/dfsm.c
+++ b/data/src/dfsm.c
@@ -162,8 +162,8 @@ static void
 dfsm_send_sync_message_abort(dfsm_t *dfsm)
 {
 	g_return_if_fail(dfsm != NULL);
-
 	g_mutex_lock (&dfsm->sync_mutex);
+	cfs_dom_debug(dfsm->log_domain, "dfsm_send_sync_message_abort - %" PRIu64" / %" PRIu64, dfsm->msgcount_rcvd, dfsm->msgcount);
 	dfsm->msgcount_rcvd = dfsm->msgcount;
 	g_cond_broadcast (&dfsm->sync_cond);
 	g_mutex_unlock (&dfsm->sync_mutex);
@@ -181,6 +181,7 @@ dfsm_record_local_result(
 
 	g_mutex_lock (&dfsm->sync_mutex);
 	dfsm_result_t *rp = (dfsm_result_t *)g_hash_table_lookup(dfsm->results, &msg_count);
+	cfs_dom_debug(dfsm->log_domain, "dfsm_record_local_result - %" PRIu64":  %d", msg_count, msg_result);
 	if (rp) {
 		rp->result = msg_result;
 		rp->processed = processed;
@@ -235,6 +236,8 @@ dfsm_send_state_message_full(
 	g_return_val_if_fail(DFSM_VALID_STATE_MESSAGE(type), CS_ERR_INVALID_PARAM);
 	g_return_val_if_fail(!len || iov != NULL, CS_ERR_INVALID_PARAM);
 
+	cfs_dom_debug(dfsm->log_domain, "dfsm_send_state_message_full: type %d len %d", type, len);
+
 	dfsm_message_state_header_t header;
 	header.base.type = type;
 	header.base.subtype = 0;
@@ -317,6 +320,7 @@ dfsm_send_message_sync(
 	for (int i = 0; i < len; i++)
 		real_iov[i + 1] = iov[i];
 
+	cfs_dom_debug(dfsm->log_domain, "dfsm_send_messag_sync: type NORMAL, msgtype %d, len %d", msgtype, len);
 	cs_error_t result = dfsm_send_message_full(dfsm, real_iov, len + 1, 1);
 
 	g_mutex_unlock (&dfsm->sync_mutex);
@@ -335,10 +339,12 @@ dfsm_send_message_sync(
 	if (rp) {
 		g_mutex_lock (&dfsm->sync_mutex);
 
-		while (dfsm->msgcount_rcvd < msgcount)
+		while (dfsm->msgcount_rcvd < msgcount) {
+			cfs_dom_debug(dfsm->log_domain, "dfsm_send_message_sync: waiting for received messages %" PRIu64 " / %" PRIu64, dfsm->msgcount_rcvd, msgcount);
 			g_cond_wait (&dfsm->sync_cond, &dfsm->sync_mutex);
+		}
+		cfs_dom_debug(dfsm->log_domain, "dfsm_send_message_sync: done waiting for received messages!");
 
-      
 		g_hash_table_remove(dfsm->results, &rp->msgcount);
 		
 		g_mutex_unlock (&dfsm->sync_mutex);
@@ -685,9 +691,13 @@ dfsm_cpg_deliver_callback(
 		return;
 	}
 
+	cfs_dom_debug(dfsm->log_domain, "received message's header type is %d", base_header->type);
+
 	if (base_header->type == DFSM_MESSAGE_NORMAL) {
 
 		dfsm_message_normal_header_t *header = (dfsm_message_normal_header_t *)msg;
+		cfs_dom_debug(dfsm->log_domain, "received normal message (type = %d, subtype = %d, %zd bytes)",
+			      base_header->type, base_header->subtype, msg_len);
 
 		if (msg_len < sizeof(dfsm_message_normal_header_t)) {
 			cfs_dom_critical(dfsm->log_domain, "received short message (type = %d, subtype = %d, %zd bytes)",
@@ -704,6 +714,8 @@ dfsm_cpg_deliver_callback(
 		} else {
 
 			int msg_res = -1;
+			cfs_dom_debug(dfsm->log_domain, "deliver message %" PRIu64 " (subtype = %d, length = %zd)",
+				      header->count, base_header->subtype, msg_len); 
 			int res = dfsm->dfsm_callbacks->dfsm_deliver_fn(
 				dfsm, dfsm->data, &msg_res, nodeid, pid, base_header->subtype, 
 				base_header->time, (uint8_t *)msg + sizeof(dfsm_message_normal_header_t),
@@ -724,6 +736,8 @@ dfsm_cpg_deliver_callback(
 	 */
 
 	dfsm_message_state_header_t *header = (dfsm_message_state_header_t *)msg;
+	cfs_dom_debug(dfsm->log_domain, "received state message (type = %d, subtype = %d, %zd bytes), mode is %d",
+			 base_header->type, base_header->subtype, msg_len, mode);
 
 	if (msg_len < sizeof(dfsm_message_state_header_t)) {
 		cfs_dom_critical(dfsm->log_domain, "received short state message (type = %d, subtype = %d, %zd bytes)",
@@ -744,6 +758,7 @@ dfsm_cpg_deliver_callback(
 
 	if (mode == DFSM_MODE_SYNCED) {
 		if (base_header->type == DFSM_MESSAGE_UPDATE_COMPLETE) {
+			cfs_dom_debug(dfsm->log_domain, "received update complete message");
 
 			for (int i = 0; i < dfsm->sync_info->node_count; i++)
 				dfsm->sync_info->nodes[i].synced = 1;
@@ -754,6 +769,7 @@ dfsm_cpg_deliver_callback(
 			return;
 
 		} else if (base_header->type == DFSM_MESSAGE_VERIFY_REQUEST) {
+			cfs_dom_debug(dfsm->log_domain, "received verify request message");
 
 			if (msg_len != sizeof(dfsm->csum_counter)) {
 				cfs_dom_critical(dfsm->log_domain, "cpg received verify request with wrong length (%zd bytes) form node %d/%d", msg_len, nodeid, pid);
@@ -823,7 +839,6 @@ dfsm_cpg_deliver_callback(
 	} else if (mode == DFSM_MODE_START_SYNC) {
 
 		if (base_header->type == DFSM_MESSAGE_SYNC_START) {
-
 			if (nodeid != dfsm->lowest_nodeid) {
 				cfs_dom_critical(dfsm->log_domain, "ignore sync request from wrong member %d/%d",
 						 nodeid, pid);
@@ -861,6 +876,7 @@ dfsm_cpg_deliver_callback(
 			return;
 
 		} else if (base_header->type == DFSM_MESSAGE_STATE) {
+			cfs_dom_debug(dfsm->log_domain, "received state message for %d/%d", nodeid, pid);
 
 			dfsm_node_info_t *ni;
 			
@@ -906,6 +922,8 @@ dfsm_cpg_deliver_callback(
 						goto leave;
 				}
 
+			} else {
+				cfs_dom_debug(dfsm->log_domain, "haven't received all states, waiting for more");
 			}
 
 			return;
@@ -914,6 +932,7 @@ dfsm_cpg_deliver_callback(
 	} else if (mode == DFSM_MODE_UPDATE) {
 
 		if (base_header->type == DFSM_MESSAGE_UPDATE) {
+			cfs_dom_debug(dfsm->log_domain, "received update message");
 				
 			int res = dfsm->dfsm_callbacks->dfsm_process_update_fn(
 				dfsm, dfsm->data, dfsm->sync_info, nodeid, pid, msg, msg_len);
@@ -925,6 +944,7 @@ dfsm_cpg_deliver_callback(
 
 		} else if (base_header->type == DFSM_MESSAGE_UPDATE_COMPLETE) {
 
+			cfs_dom_debug(dfsm->log_domain, "received update complete message");
 
 			int res = dfsm->dfsm_callbacks->dfsm_commit_fn(dfsm, dfsm->data, dfsm->sync_info);
 
@@ -1104,6 +1124,7 @@ dfsm_cpg_confchg_callback(
 	size_t joined_list_entries)
 {
 	cs_error_t result;
+	cfs_debug("dfsm_cpg_confchg_callback called");
 
 	dfsm_t *dfsm = NULL;
 	result = cpg_context_get(handle, (gpointer *)&dfsm);





More information about the pve-devel mailing list