free up not used space on a qcow2-image-file on kvm/qemu

virt-sparsify can do all this with less hassle on your part: http://libguestfs.org/virt-sparsify.1.html

The image will not shrink automatically, since when you delete files, you don't actually delete data (this is why undelete works). Qemu has a facility to shrink qcow2 images back, but what the utility does is really deduplicate the zeroes from the disk, leaving all other information intact. So the idea would be to:

1. Zero-fill the drive (dd if=/dev/zero of=/some/file until you run out of space)
2. delete /some/file
3. shut down the VM
4. cd to where the images for the VM are kept and run qemu-img convert -O qcow2 original_image.qcow2 deduplicated_image.qcow2
5. change the VM settings to use the new deduplicated_image.qcow2, test the VM is working, and remove the old image

This, afaik, will only work with qcow2 images, I haven't tested other formats.

You can use virt-sparsify for this. It automatically zeros and sparsifies free space on most filesystems it finds within the source disk image (supports ext2/3/4, btrfs, NTFS, etc) .

Step-by-step instructions for in-place conversion

1. Shut down the VM

2. Keep a backup

3. Convert: virt-sparsify --in-place disk.img

Step-by-step instructions for conversion to another file (this is safer but requires more free space):

1. Shut down the VM

2. Convert: virt-sparsify disk.img new-file.img

3. Swap old with new image file: mv disk.img disk.img.bak && mv new-file.img disk.img

4. If VM starts normally you can remove the backup: rm disk.img.bak

P.S.: You can also convert between formats when you use virt-sparsify:

virt-sparsify disk.raw --convert qcow2 disk.qcow2

In addition to the accepted answer (which describe the more common, general method to shrink a qcow2 file), modern version (ie: RHEL7+) of the libvirt/qemu/qcow2 stack supports the virtio-scsi driver which supports the discard='unmap' option.

If the virtual machine uses virtio-scsi and the libvirt definition include discard='unmap', a simple fstrim on the guest filesystem will release assigned-but-unused free space on the host. This can be checked with du -hs <guestdisk> on host side (note: ls -al <guestdisk> will not show the real allocated size, just the logical one).

For even newer libvirt/qemu instances (ie: RHEL8+), guest file size can be reduced even if the guest OS does not support trim: by enabling both discard='unmap' and discard_zeroes='unmap' (and using the virtio-scsi driver), writing a sequential stream of zero on guest side (ie: via dd if=/dev/zero of=/zero.img bs=1M count=1024) will trigger host-size trimming of the affected LBA ranges. However, please note that this can be CPU-intensive for the host (which had to "parse" any guest write searching for repeating zeroes), so it should be only enabled in specific cases.

Both methods will reduce the physical allocated size (what du -hs shows). For reducing the logical size (what a simple ls -al gives) you need to use virt-spasify or qemu-img (as described in the other answers).

virt-sparsify --compress fileA fileB will reduce the size of the image, however it creates the qcow2 image as another file, so you will have to rename and move things around. You cannot use --compress and --in-place at the same time. See the following link to learn more. http://libguestfs.org/virt-sparsify.1.html