[lxc-devel] Howto user namespaces?

[richard -rw- weinberger]

On Tue, Apr 9, 2013 at 11:07 PM, Serge Hallyn <serge.hallyn at ubuntu.com> wrote:
> Quoting richard -rw- weinberger (richard.weinberger at gmail.com):
>> On Tue, Apr 9, 2013 at 3:19 PM, Serge Hallyn <serge.hallyn at ubuntu.com> wrote:
>> > Quoting richard -rw- weinberger (richard.weinberger at gmail.com):
>> >> On Tue, Apr 9, 2013 at 9:58 AM, richard -rw- weinberger
>> >> <richard.weinberger at gmail.com> wrote:
>> >> > On Tue, Apr 9, 2013 at 5:28 AM, Serge Hallyn <serge.hallyn at ubuntu.com> wrote:
>> >> >> Quoting richard -rw- weinberger (richard.weinberger at gmail.com):
>> >> >>> Am I missing something obvious?
>> >> >>
>> >> >> lxc-create does not yet convert the rootfs to the mapped uids, so you
>> >> >> need to do that manually using uidmapshift.  Check the
>> >> >> container-userns-convert script at
>> >> >> https://code.launchpad.net/~serge-hallyn/+junk/nsexec or in the nsexec
>> >> >> package at ppa:serge-hallyn/userns-natty.
>> >> >
>> >> > Hmm, I've fixed the uids already by hand.
>> >> > Today I've created a new container and used container-userns-convert
>> >> > but with the same results.
>> >> >
>> >> > What I find very strange is that your script does:
>> >> > lxc.id_map = U ${uid} 0 $range
>> >> > lxc.id_map = G ${uid} 0 $range
>> >> > uid is 100000, range is 10000.
>> >> >
>> >> > But the lxc docs say:
>> >> >               Four values must be provided.  First a character, either
>> >> >               'u', or 'g', to specify whether user or group ids are
>> >> >               being mapped.  Next is the first userid as seen in the
>> >> >               user namespace of the container.  Next is the userid as
>> >> >               seen on the host.  Finally, a range indicating the number
>> >> >               of consecutive ids to map.
>> >> >
>> >> > So, this would make more sense: lxc.id_map = u 0 100000 10000
>> >> >
>> >> > Anyways, mount of tmpfs fails with ENOPERM, is there any debugging
>> >> > mechanism to find out why it is failing?
>> >> > According to strace some bind mounts before the tmpfs work perfectly fine.
>> >>
>> >> BTW: I found out that tmpfs is not supported within user namespaces...
>> >
>> > It should be in 3.9:
>> >
>> >     userns: Allow the userns root to mount tmpfs.
>>
>> Okay. Mounting tmpfs works on 3.9 so far.
>>
>> >> Anyways, now lxc-start dies here:
>> >> lxc-start: Operation not permitted - failed to set mode '020644' to '/dev/pts/1'
>> >> which is:
>> >> chmod("/dev/pts/1", 020644) = -1 EPERM (Operation not permitted)
>> >> Shouldn't this be /usr/lib64/lxc/rootfs/dev/pts/1?!
>> >
>> > Look at Eric's user namespaces kernel tree for patches which aren't in
>> > your tree yet.  (I also have one, but right now it is out of date with
>> > respect to some recent fixes Eric has sent upstream.)  Upstream is
>> > almost 100% there, but an <eensie weensie> bit away.
>>
>> Which tree is that?
>> git://git.kernel.org/pub/scm/linux/kernel/git/ebiederm/user-namespace.git
>> for-linus (and for-next)
>> differ only by one commit:
>> proc: Restrict mounting the proc filesystem
>
> you'll probably have better luck with branch userns-always-map-user-v100

Nope. Same result.
FWIW to build userns-always-map-user-v100 on current Linus tree the
following patch is needed:
---
diff --git a/fs/devpts/inode.c b/fs/devpts/inode.c
index f0f0faa..ecfd846 100644
--- a/fs/devpts/inode.c
+++ b/fs/devpts/inode.c
@@ -386,8 +386,8 @@ struct vfsmount *devpts_mntget(struct file *filp)
        struct vfsmount *mnt = ERR_PTR(-ENODEV);

        /* Verify the mnt is a devpts mount */
-       if (filp->f_vfsmnt->mnt_sb->s_magic == DEVPTS_SUPER_MAGIC)
-               mnt = mntget(filp->f_vfsmnt);
+       if (filp->f_path.mnt->mnt_sb->s_magic == DEVPTS_SUPER_MAGIC)
+               mnt = mntget(filp->f_path.mnt);
        return mnt;
 }
---

> The unsafe kernel (until I get time to update it) which definately works
> is in ppa ubuntu-lxc/kernel.

Okay. I'll have a look at it tomorrow.

--
Thanks,
//richard