User-Space Transformations

Supporting user-defined transformations provides substantial flexibility. However, arbitrary user code cannot be used in kernel transformations for two reasons. First, the kernel operates in privileged mode. Second, a kernel thread is not preempted; therefore, a kernel transformation cannot execute for long periods of time. As a result, arbitrary user code cannot be downloaded into the kernel.

Therefore, Mona supports user-level transformations, which, although higher in overhead, have several advantages over kernel transformations. First, a user-level transformation does not execute in supervisor mode, which preserves the integrity of the system. Second, a user-level transformation has access to all the standard programming libraries. Lastly, a user-level transformation can use an arbitrarily large amount of resources, including CPU time and memory space. These advantages combine to make a user-level transformation much simpler and more powerful than a kernel transformations.

In addition, supporting both kernel and user-level transformations makes developing kernel transformations easier. Transformation code can be developed and debugged in user space first and then inserted into the kernel.¹ This saves time because the compile/test/debug cycle is greatly reduced. First, there is no need to insert and remove a kernel module (and mount and unmount file systems) because a user transformation is an ordinary relocatable object file. Second, a symbolic debugger and print statements can be used to debug a user-level transformation. Third, a user-level bug causes a core dump, whereas a kernel bug causes a kernel panic. Developing in user space does not obviate the need for testing a transformation in the kernel, but it does reduce the time spent debugging.

  

Figure 1: Illustration of export transformation.

The Mona file system provides the export transformation to safely execute untrusted or computationally expensive transformations outside of the kernel in user space. Figure 1 illustrates a transformation network generated by connecting a combination of kernel and user-level transformations. A and B are kernel transformations, but $\alpha$ is a user-level transformation. Therefore, the file system inserts the export transformation between A and B. This transformation passes its input data across the kernel-user boundary to a user-level transformation network that runs in user space. When transformation $\alpha$ completes, the output data returns to the export transformation which then passes it on to transformation B. Section 3.2 describes the mechanics of this process in detail.