Note to self: Don't forget Qemu's discard option
If one likes their disk space, at least…
This is just a short note to myself, and to anyone who might run VMs via home-grown scripts (or systemd units). I expect modern VM managers to do this automatically, but for myself, I have just a few hacked together scripts.
By default, QEMU (at least as of version 7.0) does not honour/pass
discard requests from block devices to the underlying storage. This is
a sane default (like
lvm’s default setting), but with long-lived VMs
it can lead to lots of wasted disk space. I keep my VMs on SSDs, which
is limited space for me, so savings here are important.
Older Debian versions did not trim automatically, but nowadays they do (which is why this is worth enabling for all VMs), so all you need is to pass:
unmap=discardto activate the pass-through.
detect-zeroes=unmap, but I don’t know how useful this is, as in, how often zeroes are written.
And the next trim should save lots of disk space. It doesn’t matter
much if you use
qcow2, both will know to unmap the unused
disk, leading to less disk space used. This part seems to me safe
security-wise, as long as you trust the host. If you have pass-through
to the actual hardware, it will also do proper discard at the SSD
level (with the potential security issues leading from that). I’m
happy with the freed up disk space 🙂
Note: If you have (like I do) Windows VMs as well, using paravirt block devices, make sure the drive is recent enough.
One interesting behaviour from Windows: it looks like the default cluster size is quite high (64K), which with many small files will lead to significant overhead. But, either I misunderstand, or Windows actually knows how to unmap the unused part of a cluster (although it takes a while). So in the end, the backing file for the VM (19G) is smaller than the “disk used” as reported in Windows (23-24G), but higher than “size on disk” for all the files (17.2G). Seems legit, and it still boots 😛 Most Linux file systems have much smaller block sizes (usually 4K), so this is not a problem for it.