I have a question about a potential race, on SMP machines,
between the creat(2) and rename(2) operations.  Rename has the
semantic that if the "newpath" exists, then it will be atomically
replaced.  Now I have two processes: one is renaming a directory
"a" to directory "b" (which _is_ empty), while the second is
trying to create a file "c" within directory "b."  In that test, the
directory "b" should _always_ exist, therefore, the create should
always succeed.  However, there exists a window of opportunity
in which the create can return ENOENT.
  Upon further investigation of the VFS code, I found:
     do_rename():
        ...
        lock_kernel();
        error = vfs_rename(old_dir->d_inode, old_dentry,
                                   new_dir->d_inode, new_dentry);
        unlock_kernel();
        ...
  that eventually calls down to:
      vfs_rename_dir():
        if (target) { /* Hastur! Hastur! Hastur! */
                triple_down(&old_dir->i_zombie,
                            &new_dir->i_zombie,
                            &target->i_zombie);
                d_unhash(new_dentry);
        } else
                double_down(&old_dir->i_zombie,
                            &new_dir->i_zombie);
        ...
        if (target) {
                if (!error)
                        target->i_flags |= S_DEAD;
                triple_up(&old_dir->i_zombie,
                          &new_dir->i_zombie,
                          &target->i_zombie);
                if (d_unhashed(new_dentry))
                        d_rehash(new_dentry);
                dput(new_dentry);
        } else
                double_up(&old_dir->i_zombie,
                          &new_dir->i_zombie);
        if (!error)
                d_move(old_dentry,new_dentry);
         ...
     vfs_create():
         ...
        down(&dir->i_zombie);
        error = may_create(dir, dentry);
        if (error)
                goto exit_locks;
        ...
        lock_kernel();
        error = dir->i_op->create(dir, dentry, mode);
        unlock_kernel();
        ...
Now, in this case, both processes have performed their
path walks, and are now referencing the same dentry/inode
corresponding to the directory "b."  The rename process is
able to acquire the BKL, and the necessary i_zombies first.
The create process must then block, within vfs_create,
waiting for directory "b's" i_zombie.  The rename operation
completes, sets S_DEAD within the target inode (in this case
the target inode is the directory "b"), and releases all of its
i_zombies.  At this point, before vfs_rename_dir is able to
call d_move, the process blocked in vfs_create is able to
acquire the i_zombie, and call may_create - which finds that
the directory "b's" inode is now dead and returns ENOENT.
The problem boils down to the fact that the d_move was not
called before the may_create.  I am not sure of the appropriate
fix for this race, but I see at least two options.  One is to move
the call to may_create within the lock_kernel() - however, that
may not be a good long term solution if the BKL is going away.  
Another possibility is to move the call to d_move within the
protection of the i_zombie (within vfs_rename_dir).
The test was run on an SMP machine running Linux kernel2.4.7.  
Upon inspection, I found that the race still exists in the latest stable
release: 2.4.16.  And this race _only_ exists_ on SMP machines.
Comments?  Ideas?  Flames?
Thanks,
-Mike Gaughen
ps.  Please CC me on any replies, as I am not subscribed to the list.
-
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/