Yet another solution to the mfs unmount/kill race (not ugly this time).

Require that the mount point is vfs_busy on entry to dounmount.

1 files changed, 4 insertions, 37 deletions

diff --git a/sys/ufs/mfs/mfs_vfsops.c b/sys/ufs/mfs/mfs_vfsops.c
index 36989eab1fb..9f784417eb0 100644
--- a/sys/ufs/mfs/mfs_vfsops.c
+++ b/sys/ufs/mfs/mfs_vfsops.c
@@ -1,4 +1,4 @@
-/*	$OpenBSD: mfs_vfsops.c,v 1.10 1999/10/15 15:16:13 art Exp $	*/
+/*	$OpenBSD: mfs_vfsops.c,v 1.11 1999/12/06 07:28:06 art Exp $	*/
 /*	$NetBSD: mfs_vfsops.c,v 1.10 1996/02/09 22:31:28 christos Exp $	*/
 
 /*
@@ -66,9 +66,6 @@ static	int mfs_minor;	/* used for building internal dev_t */
 
 extern int (**mfs_vnodeop_p) __P((void *));
 
-int	mfs_dounmount __P((struct mount *));
-void	mfs_dounmount1 __P((void *));
-
 /*
  * mfs vfs operations.
  */
@@ -276,8 +273,9 @@ mfs_start(mp, flags, p)
 		 * EINTR/ERESTART.
 		 */
 		if (tsleep((caddr_t)vp, mfs_pri, "mfsidl", 0)) {
-			mfs_dounmount(mp);
-			CLRSIG(p, CURSIG(p));
+			if (vfs_busy(mp, LK_NOWAIT, NULL, p) ||
+			    dounmount(mp, 0, p))
+				CLRSIG(p, CURSIG(p));
 		}
 	}
 	return (0);
@@ -301,34 +299,3 @@ mfs_statfs(mp, sbp, p)
 		    &sbp->mount_info.mfs_args, sizeof(struct mfs_args));
 	return (error);
 }
-
-/*
- * Spawn off a kernel thread to do the unmounting to avoid all deadlocks.
- * XXX - this is horrible, but it was the only sane thing to do.
- */
-int
-mfs_dounmount(mp)
-	struct mount *mp;
-{
-	if (kthread_create(mfs_dounmount1, (void *)mp, NULL, "mfs_unmount"))
-		return 1;
-
-	return 0;
-}
-
-void
-mfs_dounmount1(v)
-	void *v;
-{
-	struct mount *mp = v;
-
-	/*
-	 * Don't try to do the unmount if someone else is trying to do that.
-	 * XXX - should be done with vfs_busy, but the problem is that
-	 *       we can't pass a locked mp into dounmount.
-	 */
-	if (!(mp->mnt_flag & MNT_UNMOUNT))
-		dounmount(mp, 0, curproc);
-
-	kthread_exit(0);
-}