Attachment 144741 Details for Bug 201396 – Makes clusvcadm / rgmanager produce an error if the node dies while handling a req.

[patch] Makes clusvcadm / rgmanager produce an error if the node dies while handling a req.

rgmanager-rhel4-nodedeath.patch (text/plain), 5.17 KB, created by Lon Hohberger on 2007-01-03 21:04:57 UTC

(hide)

Description:

Filename:

MIME Type:

Creator: Lon Hohberger

Created: 2007-01-03 21:04:57 UTC

Size: 5.17 KB

patch

obsolete

>### Eclipse Workspace Patch 1.0
>#P cluster-RHEL4
>Index: rgmanager/src/utils/clusvcadm.c
>===================================================================
>RCS file: /cvs/cluster/cluster/rgmanager/src/utils/clusvcadm.c,v
>retrieving revision 1.2.2.7
>diff -u -r1.2.2.7 clusvcadm.c
>--- rgmanager/src/utils/clusvcadm.c	13 Dec 2006 18:19:56 -0000	1.2.2.7
>+++ rgmanager/src/utils/clusvcadm.c	3 Jan 2007 20:59:10 -0000
>@@ -147,6 +147,43 @@
> }
> 
> 
>+int
>+do_msg_receive(uint64_t msgtarget, int fd, void *buf, size_t len)
>+{
>+	int ret;
>+	cluster_member_list_t *m = NULL;
>+	
>+	if ((int64_t)msgtarget < (int64_t)0)
>+		return msg_receive(fd, buf, len);
>+	
>+	/* Make sure a node hasn't died while processing our request. */
>+	do {
>+		ret = msg_receive_timeout(fd, buf, len, 10);
>+		if (ret < (int)len) {
>+			if (ret < 0 && errno == ETIMEDOUT) {
>+				m = clu_member_list(RG_SERVICE_GROUP);
>+				if (!memb_online(m, msgtarget)) {
>+					ret = RG_ENODEDEATH;
>+					break;
>+				}
>+				cml_free(m);
>+				m = NULL;
>+				continue;
>+			}
>+			
>+			/* Make sure we don't overwrite ENODEDEATH */
>+			if (ret < 0)
>+				ret = -1;
>+		}
>+		break;
>+	} while(1);
>+	
>+	if (m)
>+		cml_free(m);
>+	return ret;
>+}
>+
>+
> void
> usage(char *name)
> {
>@@ -259,7 +296,6 @@
> 		usage(basename(argv[0]));
> 		return 1;
> 	}
>-	
> 
> 	/* No login */
> 	fd = clu_connect(RG_SERVICE_GROUP, 0);
>@@ -294,10 +330,15 @@
> 		fflush(stdout);
> 		msgfd = msg_open(msgtarget, RG_PORT, 0, 5);
> 	} else {
>-		printf("Trying to relocate %s to %s", svcname, nodename);
>+		if (node_specified)
>+			printf("Trying to relocate %s to %s", svcname, nodename);
>+		else
>+			printf("Trying to relocate %s", svcname);
> 		printf("...");
> 		fflush(stdout);
> 		msgfd = msg_open(me, RG_PORT, 0, 5);
>+		/* just do a normal receive from the local node */
>+		msgtarget = (uint64_t)-1;
> 	}
> 
> 	if (msgfd < 0) {
>@@ -312,10 +353,25 @@
> 		return 1;
> 	}
> 
>-	if (msg_receive(msgfd, &msg, sizeof(msg)) != sizeof(msg)) {
>-		perror("msg_receive");
>-		fprintf(stderr, "Error receiving reply!\n");
>-		return 1;
>+	/* reusing opt */
>+	opt = do_msg_receive(msgtarget, msgfd, &msg,
>+			     sizeof(msg));
>+	if (opt < (int)sizeof(msg)) {
>+		if (opt != RG_ENODEDEATH) {
>+			perror("msg_receive");
>+			fprintf(stderr, "Error receiving reply!\n");
>+			return 1;
>+		}
>+		
>+		/*
>+		 * XXX hack to enable node death processing along side
>+		 * all the rest of the possible responses.  If an end-node 
>+		 * died while processing, this will have been set by the 
>+		 * rgmanager and a response with RG_ENODEDEATH as the d_ret
>+		 * would have been received.
>+		 */
>+		msg.sm_data.d_ret = RG_ENODEDEATH;
>+		swab_SmMessageSt(&msg);
> 	}
> 	
> 	/* Decode */
>@@ -346,6 +402,10 @@
> 	case RG_EFAIL:
> 		printf("failed\n");
> 		break;
>+	case RG_ENODEDEATH:
>+		printf("node processing request died\n");
>+		printf("(Status unknown)\n");
>+		break;
> 	case RG_EABORT:
> 		printf("cancelled by resource manager\n");
> 		break;
>Index: rgmanager/include/resgroup.h
>===================================================================
>RCS file: /cvs/cluster/cluster/rgmanager/include/resgroup.h,v
>retrieving revision 1.3.2.8
>diff -u -r1.3.2.8 resgroup.h
>--- rgmanager/include/resgroup.h	13 Dec 2006 18:19:57 -0000	1.3.2.8
>+++ rgmanager/include/resgroup.h	3 Jan 2007 20:59:10 -0000
>@@ -156,6 +156,7 @@
> cluster_member_list_t *member_list(void);
> uint64_t my_id(void);
> 
>+#define RG_ENODEDEATH	-8 /* Processing node died */
> #define RG_ERUN		-7 /* Service is running already */
> #define RG_EAGAIN	-6 /* Try again */
> #define RG_EDEADLCK	-5 /* Operation would cause deadlock */
>Index: rgmanager/src/daemons/rg_forward.c
>===================================================================
>RCS file: /cvs/cluster/cluster/rgmanager/src/daemons/rg_forward.c,v
>retrieving revision 1.2.2.2
>diff -u -r1.2.2.2 rg_forward.c
>--- rgmanager/src/daemons/rg_forward.c	13 Dec 2006 18:19:56 -0000	1.2.2.2
>+++ rgmanager/src/daemons/rg_forward.c	3 Jan 2007 20:59:10 -0000
>@@ -48,8 +48,9 @@
> 	rg_state_t rgs;
> 	request_t *req = (request_t *)arg;
> 	void *lockp;
>-	int fd;
>+	int fd, ret;
> 	SmMessageSt msg;
>+	cluster_member_list_t *m = NULL;
> 
> 	if (rg_lock(req->rr_group, &lockp) != 0) {
> 		msg_close(req->rr_resp_fd);
>@@ -88,19 +89,43 @@
> 		pthread_exit(NULL);
> 	}
> 
>-	if (msg_receive(fd, &msg, sizeof(msg)) != sizeof(msg)) {
>-		msg_close(fd);
>-		msg_close(req->rr_resp_fd);
>-		rq_free(req);
>-		pthread_exit(NULL);
>-	}
>+	/*
>+	 * Ok, we're forwarding a message to another node.  Keep tabs on 
>+	 * the node to make sure it doesn't die.  Basically, wake up every
>+	 * now and again to make sure it's still online.  If it isn't, send
>+	 * a response back to the caller.
>+	 */
>+	do {
>+		ret = msg_receive_timeout(fd, &msg, sizeof(msg), 10);
>+		if (ret < (int)sizeof(msg)) {
>+			if (ret < 0 && errno == ETIMEDOUT) {
>+				m = member_list();
>+				if (!memb_online(m, rgs.rs_owner)) {
>+					msg.sm_data.d_ret = RG_ENODEDEATH;
>+					/* we decode down below,
>+					 * so encode here */
>+					swab_SmMessageSt(&msg);
>+					break;
>+				}
>+				cml_free(m);
>+				m = NULL;
>+				continue;
>+			}
>+			msg_close(fd);
>+			msg_close(req->rr_resp_fd);
>+			goto out;
>+		}
>+		break;
>+	} while(1);
>+
>+	if (m)
>+		cml_free(m);
> 	msg_close(fd);
> 
> 	swab_SmMessageSt(&msg);
> 	send_response(msg.sm_data.d_ret, req->rr_target, req);
>-
>+out:
> 	rq_free(req);
>-
> 	pthread_exit(NULL);
> }
>

Actions: View | Diff

Attachments on bug 201396: 144739 | 144741