Nova LXD driver contains a disk usage resource leak.
Each disk image import from Glance leads to O(image_size) disk space lost in /tmp.
Leak exists until nova-compute service is restarted as during restart open descriptors will be freed and disk space will reclaimed by OS.
The reason is wrong usage of mkstemp in nova/virt/lxd/driver.py _sync_glance_image_to_lxd.
The code use mkstemp to generate "random" filename to pass to glance API as destination path for downloaded image. But mkstemp also open file description to this file which is lost and never closed. As result even when this temporary file (downloaded image) is deleted file is still locked by nova-compute process.
def _sync_glance_image_to_lxd(client, context, image_ref):
...
try:
image_file = tempfile.mkstemp()[1]
# ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
# in code above we lost msktemp[0] of tuple which is open file handle
# so file never be 'freed' until process finished
...
IMAGE_API.download(context, image_ref, dest_path=image_file)
....
Just a simple test which tries to create instant from 'bad' image uploaded to glance
show that disk space is lost. In my case it was ~= 800Mb image
root@shura-edge1:~# lsof -P 2>/dev/null | grep nova-com | grep /tmp
nova-comp 25306 nova 10u unix 0xffff88007795c000 0t0 2128952 /tmp/tmpybLoe9/
nova-comp 25306 nova 17u REG 253,1 884226169 10489 /tmp/tmpDcl9E7 (deleted)
nova-comp 25306 nova 18u REG 253,1 0 26813 /tmp/tmpGX3y7K (deleted)
nova-comp 25306 nova 26u REG 253,1 884226169 56992 /tmp/tmpWMDoKu (deleted)
nova-comp 25306 nova 27u REG 253,1 0 57308 /tmp/tmp4eR46t (deleted)
nova-comp 25306 nova 30u REG 253,1 884226169 57334 /tmp/tmpFTMcjK (deleted)
nova-comp 25306 nova 31u REG 253,1 0 57394 /tmp/tmpfZOfB1 (deleted)