cluster: RHEL58 - rgmanager: Fix for deadlock

Ryan McCabe rmccabe at fedoraproject.org
Fri Nov 16 17:55:51 UTC 2012


Gitweb:        http://git.fedorahosted.org/git/?p=cluster.git;a=commitdiff;h=07fbe0eb3b1593270b117bbefa5ddf415952c643
Commit:        07fbe0eb3b1593270b117bbefa5ddf415952c643
Parent:        9a98cfc8e11273ca14135245132d9adc90dbd754
Author:        Ryan McCabe <rmccabe at redhat.com>
AuthorDate:    Mon Oct 1 11:28:35 2012 -0400
Committer:     Ryan McCabe <rmccabe at redhat.com>
CommitterDate: Fri Nov 16 12:55:38 2012 -0500

rgmanager: Fix for deadlock

This patch fixes a deadlock in rgmanager that could occur when a node
starts rgmanager while a service is recovering.

Resolves: rhbz#834459

Acked-by: Lon Hohberger <lhh at redhat.com>
Signed-off-by: Ryan McCabe <rmccabe at redhat.com>
---
 rgmanager/src/daemons/rg_state.c   |    1 +
 rgmanager/src/daemons/rg_thread.c  |   17 ++++++++++++++++-
 rgmanager/src/daemons/service_op.c |    1 +
 3 files changed, 18 insertions(+), 1 deletions(-)

diff --git a/rgmanager/src/daemons/rg_state.c b/rgmanager/src/daemons/rg_state.c
index a8b1e36..525543d 100644
--- a/rgmanager/src/daemons/rg_state.c
+++ b/rgmanager/src/daemons/rg_state.c
@@ -2009,6 +2009,7 @@ retry:
 			/* Deliberate */
 		case RG_EDEPEND:
 		case RG_EFAIL:
+		case RG_EDEADLCK:
 			/* Uh oh - we failed to relocate to this node.
 			   ensure that we tell the next node to start it from
 			   the 'recovering' state. */
diff --git a/rgmanager/src/daemons/rg_thread.c b/rgmanager/src/daemons/rg_thread.c
index 769ca1e..60bf0c6 100644
--- a/rgmanager/src/daemons/rg_thread.c
+++ b/rgmanager/src/daemons/rg_thread.c
@@ -23,6 +23,7 @@
 #include <rg_queue.h>
 #include <assert.h>
 #include <members.h>
+#include <clulog.h>
 
 /**
  * Resource thread list entry.
@@ -749,13 +750,27 @@ rt_enqueue_request(const char *resgroupname, int request,
 			msg_free_ctx(response_ctx);
 			break;
 		}
-		fprintf(stderr, "Failed to queue request: Would block\n");
+		pthread_mutex_unlock(resgroup->rt_queue_mutex);
+		pthread_mutex_unlock(&reslist_mutex);
+		clulog(LOG_DEBUG, "Failed to queue request: Would block\n");
 		/* EWOULDBLOCK */
+		return -1;
+	}
+
+	if (resgroup->rt_request == RG_START &&
+	    (request == RG_START_REMOTE || request == RG_START_RECOVER)) {
+		send_ret(response_ctx, resgroup->rt_name, RG_EDEADLCK,
+			request, 0);
+		msg_free_ctx(response_ctx);
 		pthread_mutex_unlock(resgroup->rt_queue_mutex);
 		pthread_mutex_unlock(&reslist_mutex);
+		clulog(LOG_DEBUG,
+			"Failed to queue %d request for %s: Would block\n",
+			request, resgroupname);
 		return -1;
 	}
 
+
 	ret = rq_queue_request(resgroup->rt_queue, resgroup->rt_name,
 			       request, 0, 0, response_ctx, 0, target,
 			       arg0, arg1);
diff --git a/rgmanager/src/daemons/service_op.c b/rgmanager/src/daemons/service_op.c
index 2ea59f9..bf0e1fc 100644
--- a/rgmanager/src/daemons/service_op.c
+++ b/rgmanager/src/daemons/service_op.c
@@ -82,6 +82,7 @@ service_op_start(char *svcName,
 			++dep;
 			continue;
 		case RG_EFAIL:
+		case RG_EDEADLCK:
 			++fail;
 			continue;
 		case RG_EABORT:


More information about the cluster-commits mailing list