Comment 23 for bug 1846427

> I think I may have had the wrong image size before because both tmpfs and
> my spare LVM volume are rather limited in size.

I also had a hard time to get my image to corrupt on tmpfs because it could not grow to its final size, it seems. Sometimes qemu ran into acutal ENOSPC but most of the time lack of space on tmpfs seemed to trigger early cleanup of unused blocks and in turn prevent corruption. Only after I increased tmpfs size again and again until my machine actually started to swap would I get the spurious corruption. Both facts would seems to support your suspicion of a race condition because qcow2_detect_metadata_preallocation() would run longer the more of the image is/was allocated.