We currently create "sparse" qcow2 images for VMs that start out empty, and grow only as needed. This is great as it allows us to avoid creating too small images that later "fill up" and have to be resized, while still not wasting space.

However, currently once diskspace has been used once we never release it, even if it is no longer used inside the guest. I.e. if you removed some files this will not shrink the disk image, although in some cases (it depends on the guest fs details) the space might be reused by new files in the guest.

It would be nice if we could somehow handle the shrinking of the image too. And related to this we might want to support general "compactification" of the images including reordering the image to be in-line with the guest disk order, as well as possibly compressing disk data.

There are various approaches here.

One interesting approach is to let the guest tell us that some disk block is not used by the filesystem anymore. This is known as the TRIM operation in ata/scsi. In qemu this can be enabled with the -device option discard_granularity. This should let the guest use the trim operation, although there might have to be some configuration of the guest to get it to use this. For instance, ext4 filesystems must be mounted with the "discard" mount option atm. With this activated unused blocks in the qcow image will be marked as unused, and may be reused without growing the image. However, it doesn't *shrink* the image.

Another approach is offline compaction. At regular times we could check the image data to see how much space is unused (in case TRIM is used as per the above) and how much fragmented the image is. When things pass over a certian treshold we just do a copy of the image to a new file using qmeu-img convert, which will compact the disk, and even optionally compress it. We could then switch to the new image and start the vm.

In case TRIM doesn't work for a guest we could use the virt-sparsify tool, which works similar to TRIM but in batch form. It parses the filesystem looking for unused space on the disk and creates a new disk containing only the data the filesystem uses. This is slighly more risky as it involves parsing the guest fs data, but works without TRIM support.

Its also possible to combine the offline models with the qcow snapshot support in order to do things "online" by creating a snapshot of a running vm image, compress that, and then shutdown the vm and apply the final delta to the new image.