On 05/06/2015 02:05 PM, Michael Schwager wrote: The key for what <source> looks like is the <disk type='...'> attribute a couple lines earlier. When type='file', then <source> lists a file='...' absolute path, and libvirt treats it as a file (which sort of works, except when querying for how much of the file is actually allocated, because it assumes that stat() alone is good enough) And here, you used type='block', so <source> must list dev='...' and give an absolute path to a block device (and here, when libvirt is computing how much of the device is allocated, it relies on more than just stat(), which can give more accurate numbers when doing things like qcow2 format atop an LVM block device). Hopefully I answered it. As a rule of thumb, if it is a block device, use type='block' (or type='lun' if you need sgio passthrough support) and <source dev='...'>); otherwise use type='file' and <source file='...'. -- Eric Blake eblake redhat com +1-919-301-3266 Libvirt virtualization library http://libvirt.org

On 05/06/2015 03:27 PM, Michael Schwager wrote: VDSM does this. You gain the advantages of qcow2 (storage snapshots, efficient zero-cluster representations) coupled with less overhead than what you would get with qcow2 on a file system (cutting out the filesystem layer) and guaranteed host resources (you know the lvm size is reserved for the guest and won't grow without your action, unlike a file system where automatic growth of a qcow2 file competes for space quotas with every other file in the file system). It's still less efficient than running raw, but as close as you can get to bare-metal speeds while still gaining benefits from the format. When using qcow2-on-lvm, you are responsible for making sure the LVM volume is large enough: often (thanks to qcow2 having efficient 0-cluster representation, and particularly true if your guest supports TRIM on clusters that its file system is not using) the guest will see more virtual space than what the actual lvm volume exposes (a guest that needs 10G of storage but isn't fully using it can sometimes get by with a 5G lvm partition, for example); although the converse is also possible (a guest that can only see 10G of storage, but where you have lots of qcow2 snapshots, may require 20G of lvm storage to track all the qcow2 metadata). Another use case for using qcow2 on lvm is that you can initially oversubscribe your storage: qemu has a mode where it will gracefully halt the guest if it runs out of space required for managing the qcow2 data, at which point you can then enlarge the lvm volume and resume the guest; the guest will never know that you were shuffling the size of the host storage (note that this applies for twp separate cases - one where you are merely giving the host more room to store qcow2 metadata, but leaving the guest size unchanged, such as when you start a guest that sees a 10G disk on an lvm partition that is initially 5G large, then enlarge the lvm partition to 6G but the guest still sees a 10G disk; the other where you are actually resizing the amount of storage allocated to the guest as if you had hot-swapped in a larger disk, and where some guests can then resize their file system on the fly to take advantage of the larger disk). Related to this, I'm currently working on a patch for libvirt that will take advantage of a recent qemu feature of setting a threshold for event triggering, such that you could then request notification that the lvm volume is X% full and give yourself time to enlarge the lvm volume even before you reach the point where the guest would halt because the lvm volume was full. -- Eric Blake eblake redhat com +1-919-301-3266 Libvirt virtualization library http://libvirt.org