Andy Lutomirski <luto@xxxxxxxxxxxxxx> writes: > On Tue, Oct 7, 2014 at 2:26 PM, Eric W. Biederman <ebiederm@xxxxxxxxxxxx> wrote: >> Andy Lutomirski <luto@xxxxxxxxxxxxxx> writes: >> >>> Why should MNT_LOCKED on submounts be enforced? >>> >>> Is it because, if you retain a reference to the detached tree, then >>> you can see under the submounts? >> >> Yes. MNT_DETACH is a recursive operation that detaches all of the mount >> and all of it's submounts. Which means you can see under the submounts >> if you have a reference to a detached mount. >> >>> If so, let's fix *that*. Because >>> otherwise the whole model of pivot_root + detach will break. >> >> I am not certain what you are referring to. pivot_root doesn't >> manipulate the mount tree so you can see under anything. >> >> What I believe is the appropriate fix is to fail umount2(...,MNT_DETACH) >> if there are any referenced mount points being detached that have a >> locked submount. > > Most of the container-using things do, roughly: > > Unshare userns and mountns > Mount some new stuff > pivot_root to the new stuff > MNT_DETACH the old. > > That last step will almost always fail if you make this change. I don't think so. I expect I could add full busy detection of normal umounts and those applications would not fail. What I am proposing is a more targeted version of busy detection that looks at each mount in the set that detach will unmount. For each mount if it is busy with non-submount references and it has at least one locked submount fail the detach with -EBUSY. Do you really think we have userspace references to the one or more of the mounts under old? Eric -- To unsubscribe from this list: send the line "unsubscribe linux-api" in the body of a message to majordomo@xxxxxxxxxxxxxxx More majordomo info at http://vger.kernel.org/majordomo-info.html
