Ok, here is the promised "consistent reproducible picture" and, btw,
the "-o noatime" observation seemed to be bogus (i.e. some small changes
I made to distorted the picture). Below is a session with a clean
2.3.99-pre7-5 (with a tiny debug patch below):

# mount -t ramfs none /mnt/m3
# cd /mnt/m3
# touch new
# l
total 0
-rw-r--r--    1 root     root            0 May  5 17:33 new
# cd
# umount /mnt/m3
C: count!=2 (3)
umount: /mnt/m3: device is busy

# mount -t ramfs -o noatime none /mnt/m4
# cd /mnt/m4
# touch new
# l
total 0
-rw-r--r--    1 root     root            0 May  5 17:34 new
# cd
# umount /mnt/m4
C: count!=2 (3)
umount: /mnt/m4: device is busy

# mount -t ramfs none /mnt/m5
# touch /mnt/m5/new
# umount /mnt/m5
C: count!=2 (3)
umount: /mnt/m5: device is busy
# rm /mnt/m5/new
# umount /mnt/m5
A: mnt_count>2 (3)
umount: /mnt/m5: device is busy
# l /mnt/m5
total 0
# umount /mnt/m5
A: mnt_count>2 (3)
umount: /mnt/m5: device is busy


--- linux/fs/super.c	Fri May  5 08:20:28 2000
+++ work/fs/super.c	Fri May  5 17:22:30 2000
@@ -921,6 +921,8 @@
 	}
 
 	if (atomic_read(&mnt->mnt_count) > 2) {
+		printk(KERN_ERR "A: mnt_count>2 (%d)\n",
+				atomic_read(&mnt->mnt_count));
 		mntput(mnt);
 		return -EBUSY;
 	}
@@ -965,6 +967,8 @@
 	/* Something might grab it again - redo checks */
 
 	if (atomic_read(&mnt->mnt_count) > 2) {
+		printk(KERN_ERR "B: mnt_count>2 (%d)\n",
+				atomic_read(&mnt->mnt_count));
 		mntput(mnt);
 		return -EBUSY;
 	}
@@ -980,11 +984,16 @@
 	count = d_active_refs(sb->s_root);
 	if (mnt->mnt_parent == mnt)
 		count--;
-	if (count != 2)
+	if (count != 2) {
+		printk(KERN_ERR "C: count!=2 (%d)\n", count);
 		return -EBUSY;
+	}
 
-	if (sb->s_root->d_inode->i_state)
+	if (sb->s_root->d_inode->i_state) {
+		printk(KERN_ERR "D: i_state=%d\n", 
+				sb->s_root->d_inode->i_state);
 		return -EBUSY;
+	}
 
 	/* OK, that's the point of no return */
 	mntput(mnt);