[Libguestfs] Remote use of libguestfs

[Stefan asked an interesting question on IRC about how / whether it is possible to use libguestfs remotely. The answer is too detailed for IRC] Consider this model of the libguestfs architecture[i]: (2) libguestfs (3) Linux (1) API protocol syscalls program <--------> libguestfs <--------> qemu <--------> devices guestfsd disk images kernel There are three places where you might want to use libguestfs remotely. You can remote libguestfs at (3). For example if your disk is located remotely and made available to the local machine using iSCSI, it appears as /dev/sdX and you can just access that directly with libguestfs. Another possibility is to use the NBD or http drivers built into qemu. Another possibility is to use sshfs (which we already advocate for access to VMware ESX VMFS[ii]). One user has successfully remoted libguestfs at (1), but it depends on the programming language you are using. He used rpyc[iii] to automatically generate Python stubs for libguestfs plus a remote daemon. Same is probably possible in Java using (eg) RMI. It's obviously a lot more complicated if you just consider the C API. The interesting case is (2), and I believe it's also possible to remote libguestfs here. In libguestfs 1.10 we added support for accessing live guests. The way this works under the hood is that instead of launching the qemu backend, you tell libguestfs to connect to an existing Unix domain socket. To access a live guest, the idea is that this is set up to point to a virtio-serial channel into the running guest. But you could use the same underlying API[iv] to connect to a remote qemu that you run yourself. You could start the remote qemu using (eg.) ssh and have it forward the virtio-serial channel back to something on the local side (nc -U?). Additionally you would have to work out how to send the appliance over to the remote side and/or arrange that there's a compatible appliance and qemu available to run remotely. As a final note, I should add that libguestfs will have different performance characteristics depending on which remote method you use. The reason is that some functionality exists only in the library (eg. setting and getting various flags), some functionality runs mainly in the library but makes lots of calls over the protocol (eg. the core inspection API), and some functionality happens almost entirely in guestfsd. Rich. [i] http://libguestfs.org/guestfs.3.html#architecture [ii] https://rwmj.wordpress.com/2011/05/10/tip-use-libguestfs-on-vmware-esx-gu... [iii] http://rpyc.wikidot.com/ [iv] http://libguestfs.org/guestfs.3.html#attaching_to_running_daemons -- Richard Jones, Virtualization Group, Red Hat http://people.redhat.com/~rjones New in Fedora 11: Fedora Windows cross-compiler. Compile Windows programs, test, and build Windows installers. Over 70 libraries supprt'd http://fedoraproject.org/wiki/MinGW http://www.annexia.org/fedora_mingw