Execute-in-place

Carsten Otte (of IBM) has posted a set of patches adding XIP support to the 2.6 kernel. These patches, in addition, enable fast memory-to-memory block I/O for such devices, shorting out the page cache and most of the block layer. As a result, the XIP patches are useful in a number of situations, such as, as Carsten notes, for shared-memory block devices used to communicate between (virtual) systems.

The first step is to add support at the block driver level. To that end, a new method is added to the block_device_operations structure:

    int (*direct_access) (struct inode *inode, sector_t sector, 
                          unsigned long *data);

This method, if implemented, should come up with a kernel virtual address corresponding to the given sector on the block device represented by inode. That address, which must remain valid until the device is closed, is returned in *data. The return value is zero on success or a negative error code in case of problems.

The next step is a new method in the address_space_operations structure:

    struct page *(*get_xip_page)(struct address_space *space, 
                                 sector_t blockno, int create);

This method's job is to translate a specific block number within a filesystem to a page structure pointing to its directly-mapped data. It is a filesystem-specific function which will translate blockno to a sector number on the underlying device, then use that device's direct_access() method to get an address. Carsten has posted an implementation for ext2 which shows how this method can be put together.

So far, the XIP patches enable fast, memory-to-memory device access, but they do not yet implement true execute-in-place operation. The last step is to replace the usual nopage() VMA operation (filemap_nopage()) with a new version (filemap_xip_nopage()) when the underlying device and filesystem support XIP. The new nopage() method will (using get_xip_page()) handle page faults by causing a process's page tables to point directly to the on-"disk" pages, rather than reading those pages into RAM. Some other technique will be needed to run the kernel itself in an XIP mode, but anything that is invoked thereafter can be run directly from the memory device.

Put the above pieces together, and Linux has a complete execute-in-place implementation. Supporting XIP at the block level is not the only way it could be implemented; David Woodhouse pointed out that an alternative approach is to use a special-purpose filesystem. Carsten's patches, however, point out a way in which any filesystem could be made to work in an XIP mode.