Details
-
Bug
-
Status: Resolved
-
Major
-
Resolution: Fixed
-
None
-
None
-
Mesosphere Sprint 26, Mesosphere Sprint 27, Mesosphere Sprint 29
-
2
Description
I noticed this when I was testing the unified containerizer with the bind mount backend and no volumes.
The current implementation of fs::enter will put the old root under /tmp/.old_root.XXXXXX in the new rootfs. It assumes that /tmp is writable in the new rootfs, but this might not be true, especially if the bind mount backend is used.
To solve the problem, what we can do is to mount tmpfs to /tmp in the new rootfs and umount it after pivot_root.