1 files changed, 56 insertions, 18 deletions

diff --git a/fs/ocfs2/dlm/userdlm.c b/fs/ocfs2/dlm/userdlm.c
index c3764f4744e..74ca4e5f976 100644
--- a/fs/ocfs2/dlm/userdlm.c
+++ b/fs/ocfs2/dlm/userdlm.c
@@ -139,6 +139,10 @@ static void user_ast(void *opaque)
 		return;
 	}
 
+	mlog_bug_on_msg(lockres->l_requested == LKM_IVMODE,
+			"Lockres %s, requested ivmode. flags 0x%x\n",
+			lockres->l_name, lockres->l_flags);
+
 	/* we're downconverting. */
 	if (lockres->l_requested < lockres->l_level) {
 		if (lockres->l_requested <=
@@ -229,23 +233,42 @@ static void user_unlock_ast(void *opaque, enum dlm_status status)
 
 	mlog(0, "UNLOCK AST called on lock %s\n", lockres->l_name);
 
-	if (status != DLM_NORMAL)
+	if (status != DLM_NORMAL && status != DLM_CANCELGRANT)
 		mlog(ML_ERROR, "Dlm returns status %d\n", status);
 
 	spin_lock(&lockres->l_lock);
-	if (lockres->l_flags & USER_LOCK_IN_TEARDOWN)
+	/* The teardown flag gets set early during the unlock process,
+	 * so test the cancel flag to make sure that this ast isn't
+	 * for a concurrent cancel. */
+	if (lockres->l_flags & USER_LOCK_IN_TEARDOWN
+	    && !(lockres->l_flags & USER_LOCK_IN_CANCEL)) {
 		lockres->l_level = LKM_IVMODE;
-	else {
+	} else if (status == DLM_CANCELGRANT) {
+		mlog(0, "Lock %s, cancel fails, flags 0x%x\n",
+		     lockres->l_name, lockres->l_flags);
+		/* We tried to cancel a convert request, but it was
+		 * already granted. Don't clear the busy flag - the
+		 * ast should've done this already. */
+		BUG_ON(!(lockres->l_flags & USER_LOCK_IN_CANCEL));
+		lockres->l_flags &= ~USER_LOCK_IN_CANCEL;
+		goto out_noclear;
+	} else {
+		BUG_ON(!(lockres->l_flags & USER_LOCK_IN_CANCEL));
+		/* Cancel succeeded, we want to re-queue */
+		mlog(0, "Lock %s, cancel succeeds, flags 0x%x\n",
+		     lockres->l_name, lockres->l_flags);
 		lockres->l_requested = LKM_IVMODE; /* cancel an
 						    * upconvert
 						    * request. */
 		lockres->l_flags &= ~USER_LOCK_IN_CANCEL;
 		/* we want the unblock thread to look at it again
 		 * now. */
-		__user_dlm_queue_lockres(lockres);
+		if (lockres->l_flags & USER_LOCK_BLOCKED)
+			__user_dlm_queue_lockres(lockres);
 	}
 
 	lockres->l_flags &= ~USER_LOCK_BUSY;
+out_noclear:
 	spin_unlock(&lockres->l_lock);
 
 	wake_up(&lockres->l_event);
@@ -268,13 +291,26 @@ static void user_dlm_unblock_lock(void *opaque)
 
 	spin_lock(&lockres->l_lock);
 
-	BUG_ON(!(lockres->l_flags & USER_LOCK_BLOCKED));
-	BUG_ON(!(lockres->l_flags & USER_LOCK_QUEUED));
+	mlog_bug_on_msg(!(lockres->l_flags & USER_LOCK_QUEUED),
+			"Lockres %s, flags 0x%x\n",
+			lockres->l_name, lockres->l_flags);
 
-	/* notice that we don't clear USER_LOCK_BLOCKED here. That's
-	 * for user_ast to do. */
+	/* notice that we don't clear USER_LOCK_BLOCKED here. If it's
+	 * set, we want user_ast clear it. */
 	lockres->l_flags &= ~USER_LOCK_QUEUED;
 
+	/* It's valid to get here and no longer be blocked - if we get
+	 * several basts in a row, we might be queued by the first
+	 * one, the unblock thread might run and clear the queued
+	 * flag, and finally we might get another bast which re-queues
+	 * us before our ast for the downconvert is called. */
+	if (!(lockres->l_flags & USER_LOCK_BLOCKED)) {
+		mlog(0, "Lockres %s, flags 0x%x: queued but not blocking\n",
+			lockres->l_name, lockres->l_flags);
+		spin_unlock(&lockres->l_lock);
+		goto drop_ref;
+	}
+
 	if (lockres->l_flags & USER_LOCK_IN_TEARDOWN) {
 		mlog(0, "lock is in teardown so we do nothing\n");
 		spin_unlock(&lockres->l_lock);
@@ -282,7 +318,9 @@ static void user_dlm_unblock_lock(void *opaque)
 	}
 
 	if (lockres->l_flags & USER_LOCK_BUSY) {
-		mlog(0, "BUSY flag detected...\n");
+		mlog(0, "Cancel lock %s, flags 0x%x\n",
+		     lockres->l_name, lockres->l_flags);
+
 		if (lockres->l_flags & USER_LOCK_IN_CANCEL) {
 			spin_unlock(&lockres->l_lock);
 			goto drop_ref;
@@ -296,14 +334,7 @@ static void user_dlm_unblock_lock(void *opaque)
 				   LKM_CANCEL,
 				   user_unlock_ast,
 				   lockres);
-		if (status == DLM_CANCELGRANT) {
-			/* If we got this, then the ast was fired
-			 * before we could cancel. We cleanup our
-			 * state, and restart the function. */
-			spin_lock(&lockres->l_lock);
-			lockres->l_flags &= ~USER_LOCK_IN_CANCEL;
-			spin_unlock(&lockres->l_lock);
-		} else if (status != DLM_NORMAL)
+		if (status != DLM_NORMAL)
 			user_log_dlm_error("dlmunlock", status, lockres);
 		goto drop_ref;
 	}
@@ -581,6 +612,14 @@ int user_dlm_destroy_lock(struct user_lock_res *lockres)
 	mlog(0, "asked to destroy %s\n", lockres->l_name);
 
 	spin_lock(&lockres->l_lock);
+	if (lockres->l_flags & USER_LOCK_IN_TEARDOWN) {
+		mlog(0, "Lock is already torn down\n");
+		spin_unlock(&lockres->l_lock);
+		return 0;
+	}
+
+	lockres->l_flags |= USER_LOCK_IN_TEARDOWN;
+
 	while (lockres->l_flags & USER_LOCK_BUSY) {
 		spin_unlock(&lockres->l_lock);
 
@@ -606,7 +645,6 @@ int user_dlm_destroy_lock(struct user_lock_res *lockres)
 
 	lockres->l_flags &= ~USER_LOCK_ATTACHED;
 	lockres->l_flags |= USER_LOCK_BUSY;
-	lockres->l_flags |= USER_LOCK_IN_TEARDOWN;
 	spin_unlock(&lockres->l_lock);
 
 	mlog(0, "unlocking lockres %s\n", lockres->l_name);