From: David Teigland <teigland@redhat.com>
Subject: [RHEL5 PATCH] dlm: don't accept replies to old recovery messages
Date: Mon, 27 Nov 2006 12:21:08 -0600
Bugzilla: 215430
Message-Id: <20061127182108.GE18093@redhat.com>
Changelog: dlm: don't accept replies to old recovery messages


bz 215430
patch 2 of 2

We often abort a recovery after sending a status request to a remote node.
We want to ignore any potential status reply we get from the remote node.
If we get one of these unwanted replies, we've often moved on to the next
recovery message and incremented the message sequence counter, so the
reply will be ignored due to the seq number.  In some cases, we've not
moved on to the next message so the seq number of the reply we want to
ignore is still correct, causing the reply to be accepted.  The next
recovery message will then mistake this old reply as a new one.

To fix this, we add the flag RCOM_WAIT to indicate when we can accept a
new reply.  We clear this flag if we abort recovery while waiting for a
reply.  Before the flag is set again (to allow new replies) we know that
any old replies will be rejected due to their sequence number.  We also
initialize the recovery-message sequence number to a random value when a
lockspace is first created.  This makes it clear when messages are being
rejected from an old instance of a lockspace that has since been
recreated.


Index: latest/fs/dlm/rcom.c
===================================================================
--- latest.orig/fs/dlm/rcom.c
+++ latest/fs/dlm/rcom.c
@@ -90,13 +90,28 @@ static int check_config(struct dlm_ls *l
 	return 0;
 }
 
+static void allow_sync_reply(struct dlm_ls *ls, uint64_t *new_seq)
+{
+	spin_lock(&ls->ls_rcom_spin);
+	*new_seq = ++ls->ls_rcom_seq;
+	set_bit(LSFL_RCOM_WAIT, &ls->ls_flags);
+	spin_unlock(&ls->ls_rcom_spin);
+}
+
+static void disallow_sync_reply(struct dlm_ls *ls)
+{
+	spin_lock(&ls->ls_rcom_spin);
+	clear_bit(LSFL_RCOM_WAIT, &ls->ls_flags);
+	clear_bit(LSFL_RCOM_READY, &ls->ls_flags);
+	spin_unlock(&ls->ls_rcom_spin);
+}
+
 int dlm_rcom_status(struct dlm_ls *ls, int nodeid)
 {
 	struct dlm_rcom *rc;
 	struct dlm_mhandle *mh;
 	int error = 0;
 
-	memset(ls->ls_recover_buf, 0, dlm_config.buffer_size);
 	ls->ls_recover_nodeid = nodeid;
 
 	if (nodeid == dlm_our_nodeid()) {
@@ -108,12 +123,14 @@ int dlm_rcom_status(struct dlm_ls *ls, i
 	error = create_rcom(ls, nodeid, DLM_RCOM_STATUS, 0, &rc, &mh);
 	if (error)
 		goto out;
-	rc->rc_id = ++ls->ls_rcom_seq;
+
+	allow_sync_reply(ls, &rc->rc_id);
+	memset(ls->ls_recover_buf, 0, dlm_config.buffer_size);
 
 	send_rcom(ls, mh, rc);
 
 	error = dlm_wait_function(ls, &rcom_response);
-	clear_bit(LSFL_RCOM_READY, &ls->ls_flags);
+	disallow_sync_reply(ls);
 	if (error)
 		goto out;
 
@@ -150,14 +167,20 @@ static void receive_rcom_status(struct d
 
 static void receive_sync_reply(struct dlm_ls *ls, struct dlm_rcom *rc_in)
 {
-	if (rc_in->rc_id != ls->ls_rcom_seq) {
-		log_debug(ls, "reject old reply %d got %llx wanted %llx",
-			  rc_in->rc_type, rc_in->rc_id, ls->ls_rcom_seq);
-		return;
+	spin_lock(&ls->ls_rcom_spin);
+	if (!test_bit(LSFL_RCOM_WAIT, &ls->ls_flags) ||
+	    rc_in->rc_id != ls->ls_rcom_seq) {
+		log_debug(ls, "reject reply %d from %d seq %llx expect %llx",
+			  rc_in->rc_type, rc_in->rc_header.h_nodeid,
+			  rc_in->rc_id, ls->ls_rcom_seq);
+		goto out;
 	}
 	memcpy(ls->ls_recover_buf, rc_in, rc_in->rc_header.h_length);
 	set_bit(LSFL_RCOM_READY, &ls->ls_flags);
+	clear_bit(LSFL_RCOM_WAIT, &ls->ls_flags);
 	wake_up(&ls->ls_wait_general);
+ out:
+	spin_unlock(&ls->ls_rcom_spin);
 }
 
 static void receive_rcom_status_reply(struct dlm_ls *ls, struct dlm_rcom *rc_in)
@@ -171,7 +194,6 @@ int dlm_rcom_names(struct dlm_ls *ls, in
 	struct dlm_mhandle *mh;
 	int error = 0, len = sizeof(struct dlm_rcom);
 
-	memset(ls->ls_recover_buf, 0, dlm_config.buffer_size);
 	ls->ls_recover_nodeid = nodeid;
 
 	if (nodeid == dlm_our_nodeid()) {
@@ -185,12 +207,14 @@ int dlm_rcom_names(struct dlm_ls *ls, in
 	if (error)
 		goto out;
 	memcpy(rc->rc_buf, last_name, last_len);
-	rc->rc_id = ++ls->ls_rcom_seq;
+
+	allow_sync_reply(ls, &rc->rc_id);
+	memset(ls->ls_recover_buf, 0, dlm_config.buffer_size);
 
 	send_rcom(ls, mh, rc);
 
 	error = dlm_wait_function(ls, &rcom_response);
-	clear_bit(LSFL_RCOM_READY, &ls->ls_flags);
+	disallow_sync_reply(ls);
  out:
 	return error;
 }
Index: latest/fs/dlm/dlm_internal.h
===================================================================
--- latest.orig/fs/dlm/dlm_internal.h
+++ latest/fs/dlm/dlm_internal.h
@@ -471,6 +471,7 @@ struct dlm_ls {
 	char			*ls_recover_buf;
 	int			ls_recover_nodeid; /* for debugging */
 	uint64_t		ls_rcom_seq;
+	spinlock_t		ls_rcom_spin;
 	struct list_head	ls_recover_list;
 	spinlock_t		ls_recover_list_lock;
 	int			ls_recover_list_count;
@@ -488,7 +489,8 @@ struct dlm_ls {
 #define LSFL_RUNNING		1
 #define LSFL_RECOVERY_STOP	2
 #define LSFL_RCOM_READY		3
-#define LSFL_UEVENT_WAIT	4
+#define LSFL_RCOM_WAIT		4
+#define LSFL_UEVENT_WAIT	5
 
 /* much of this is just saving user space pointers associated with the
    lock that we pass back to the user lib with an ast */
Index: latest/fs/dlm/lockspace.c
===================================================================
--- latest.orig/fs/dlm/lockspace.c
+++ latest/fs/dlm/lockspace.c
@@ -473,6 +473,8 @@ static int new_lockspace(char *name, int
 	ls->ls_recoverd_task = NULL;
 	mutex_init(&ls->ls_recoverd_active);
 	spin_lock_init(&ls->ls_recover_lock);
+	spin_lock_init(&ls->ls_rcom_spin);
+	get_random_bytes(&ls->ls_rcom_seq, sizeof(uint64_t));
 	ls->ls_recover_status = 0;
 	ls->ls_recover_seq = 0;
 	ls->ls_recover_args = NULL;