When to Use What

Now that you have learned about these three mechanisms for providing filesystem data to a UML, the question remains: Under what circumstances should you use each of them? The answer is fairly easy for hostfsnormally, it should be used only for access to host files that belong to the user owning the UML or to files that are available readonly. In the first case, the user should be logged in to the UML as root, and there should be no other UML users accessing the hostfs mount. In the second, the read-only restriction avoids all of the permission and ownership issues with hostfs.

humfs hierarchies and ubd images can be used to provide general-purpose filesystems, including root filesystems. humfs provides easier access to the UML files, although some care is needed when changing those files in order to ensure that the file metadata is updated properly.

There are also some potential efficiency advantages with both humfs and ubd devices. An issue with host memory consumption is that both the host and UML will generally cache file data separately. As a result, the host's memory will contain multiple copies of UML file data, one in the host's page cache and one for each UML that has read the data.

ubd devices can avoid this double caching by using O_DIRECT I/O on 2.6 hosts. O_DIRECT avoids the use of the host page cache, so the only copies of the data will be in the UMLs that have read it. In order to truly minimize host memory consumption, this should be used only for data that's private to the UML, such as a private filesystem image or a COW file. For a COW file, the memory savings obtained by avoiding the double caching are probably outweighed by the duplicate caching of the backing file data in the UMLs that are sharing it.

For shared data, humfs avoids the double caching by mapping the data from the host. The data is cached on the host, but mapping it provides the UML with the same page of memory that's in the host page cache. Taking advantage of this would require a form of COW for humfs, which currently doesn't exist. A file-level form of COW is possible and may exist by the time you read this. With this, a humfs equivalent of a backing file, in the form of a read-only host directory hierarchy, would be mapped into the UMLs that share it. They would all share the same memory, so there would be only one copy of it in the host's memory.

In short, both ubd devices and humfs directories have a place in a well-run UML installation. The use of one or the other should be driven by the importance of convenient host access to the UML filesystem, the ease and speed of making backups of the data, and avoidance of excessive host memory consumption.