Are you referring to user namespaces and, if so, how does that kind of break out to host root work? I thought the whole point of user namespaces was your UID 0 inside the container is UID 100000 or whatever from the perspective of outside the container. Escaping the container shouldn't inherently grant you ability to change your actual UID in the host's main namespace in that kind of setup, but I'm not sure Docker actually leverages user namespaces or not.
E.g. on my systemd-nspawn setup with --private-users=pick (enables user namespacing) I created a container and gave it a bind mount. From the container it appears like files in the bind mount created by the container namespace's UID 0 are owned by UID 0 but from outside the container the same file looks owned by UID 100000. Inverted, files owned by the "real" UID 0 on the host look owned by 0 to the host but as owned by 65534 (i.e. "nobody") from the container's perspective. Breaking out of the container shouldn't inherently change the "actual" user of the process from 100000 to 0 any more than breaking out of the container as a non-0 UID in the first place - same as breaking out of any of the other namespaces doesn't make the "UID 0" user in the container turn into "UID 0" on the host.
alt Hacker News
Are you referring to user namespaces and, if so, how does that kind of break out to host root work? I thought the whole point of user namespaces was your UID 0 inside the container is UID 100000 or whatever from the perspective of outside the container. Escaping the container shouldn't inherently grant you ability to change your actual UID in the host's main namespace in that kind of setup, but I'm not sure Docker actually leverages user namespaces or not.
E.g. on my systemd-nspawn setup with --private-users=pick (enables user namespacing) I created a container and gave it a bind mount. From the container it appears like files in the bind mount created by the container namespace's UID 0 are owned by UID 0 but from outside the container the same file looks owned by UID 100000. Inverted, files owned by the "real" UID 0 on the host look owned by 0 to the host but as owned by 65534 (i.e. "nobody") from the container's perspective. Breaking out of the container shouldn't inherently change the "actual" user of the process from 100000 to 0 any more than breaking out of the container as a non-0 UID in the first place - same as breaking out of any of the other namespaces doesn't make the "UID 0" user in the container turn into "UID 0" on the host.