Check out the new USENIX Web site. next up previous
Next: 3. Evaluation Up: FiST: A Language for Previous: 1. Introduction

   
2. Implementation

We implemented the FiST system in Solaris, Linux, and FreeBSD because these three operating systems span the most popular modern Unix platforms and they are sufficiently different from each other. This forced us to understand the generic problems in addition to the system-specific problems. Also, we had access to kernel sources for all three platforms, which proved valuable during the development of our templates. Finally, all three platforms support loadable kernel modules, which sped up the development and debugging process. Loadable kernel modules are a convenience in implementing FiST; they are not required.

The implementation of Basefs was simple and improved on previously reported efforts[27]. No changes were required to either Solaris or FreeBSD. No changes to Linux were required if using statically linked modules. To use dynamically loadable kernel modules under Linux, only three lines of code were changed in a header file. This change was passive and did not have any impact on the Linux kernel.

The remainder of this section describes the implementation of fistgen. Fistgen translates FiST code into C code which implements the file system described in the FiST input file. The code can be compiled as a dynamically loadable kernel module or statically linked with a kernel. In this section we describe the implementation of key features of FiST that span its full range of capabilities.

We implemented read-only execution environment variables (Section 2.3.1) such as %uid by looking for them in one of the fields from struct cred in Solaris or struct ucred in FreeBSD. The VFS passes these structures to vnode functions. The Linux VFS simplifies access to credentials by reading that information from the disk inode and into the in-memory vnode structure, struct inode. So on Linux we find UID and other credentials by referencing a field directly in the inode which the VFS passes to us.

Most of the vnode attributes listed Section 2.3.1 are simple to find. On Linux they are part of the main vnode structure. On Solaris and FreeBSD, however, we first perform a VOP_GETATTR vnode operation to find them, and then return the appropriate field from the structure that the getattr function fills.

The vnode attribute ``name'' was more complex to implement, because most kernels do not store file names after the initial name lookup routine translates the name to a vnode. On Linux, implementing the vnode name attribute was simple, because it is part of a standard directory entry structure, dentry. On Solaris and FreeBSD, however, we add code to the lookup vnode function that stores the initial file name in the private data of the vnode. That way we can access it as any other vnode attribute, or any other per-vnode attribute added using the per_vnode declaration. We implemented all other fields defined using the per_vfs FiST declaration in a similar fashion.

The FiST declarations described in Section 2.3 affect the overall behavior of the generated file system. We implemented the read-only access mode by replacing the call part of every file system function that modifies state (such as unlink and mkdir) to return the error code ``read-only file system.'' We implemented the fan-in mount style by excluding code that uses the mounted directory's vnode also as the mount point.

The only difficult part of implementing the ioctl declaration and its associated functions, fistGetIoctlData and fistSetIoctlData (Section 2.2), was finding how to copy data between user space and kernel space. Solaris and FreeBSD use the routines copyin and copyout; Linux 2.3 uses copy_from_user and copy_to_user.

The last complex feature we implemented was the fileformat FiST declaration and the functions used with it: fistGetFileData and fistSetFileData (Section 2.3.1). Consider this small code excerpt:

fileformat fmt { data structure;}
fistGetFileData(file, fmt, field, out);

First, we generate a C data structure named fmt. To implement fistGetFileData, we open file, read as many bytes from it as the size of the data structure, map these bytes onto a temporary variable of the same data structure type, copy the desired field within that data structure into out, close the file, and finally return a error/success status value from the function. To improve performance, if fileformat related functions are called several times inside a vnode function, we keep the file they refer to open until the last call that uses it.

Fistgen itself (excluding the templates) is highly portable, and can be compiled on any Unix system. The total number of source lines for fistgen is 4813. Fistgen can process each 1KB of template data in under 0.25 seconds (measured on the same platform used in Section 5.3).


next up previous
Next: 3. Evaluation Up: FiST: A Language for Previous: 1. Introduction
Erez Zadok
2000-04-25