For review: user_namespaces(7) man page
rob at landley.net
Mon Apr 29 19:45:46 UTC 2013
On 04/29/2013 02:45:45 AM, richard -rw- weinberger wrote:
> On Thu, Mar 21, 2013 at 4:52 PM, Michael Kerrisk (man-pages)
> <mtk.manpages at gmail.com> wrote:
> > Hi Serge,
> > On Fri, Mar 15, 2013 at 4:38 PM, Serge Hallyn
> <serge.hallyn at ubuntu.com> wrote:
> >> Hi,
> >> you mention that after creating a new user namespace you at first
> >> all capabilities in the new ns. You don't explicitly mention (or I
> >> missed it - I did see the mention of securebits) that if you want
> >> keep those capabilities after doing an exec, you need to first have
> >> something mapped to uid 0 in the userns, and do setuid(0).
> > Good point. I'll add something on that.
> >> You might not want to list manpages from other projects,
> > Actually, not a problem. Many of the pages in my set already do
> >> but Eric's
> >> shadow patches introduce some good new manpages as well. Those
> aren't yet
> >> accepted upstream, but if/when they are then mention at least of
> >> subuid(5), subgid(5), and newuidmap(1) and newgidmap(1) might be
> > I'll add those.
> > Cheers,
> > Michael
> Can we please also document which capabilities are useless within an
> user namespace?
> E.g. CAP_MKNOD.
> You get this capability but the kernel always checks it against the
> initial userns.
Is there a device namespace?
I vaguely recall OpenVZ implemented moving individual PCI devices into
containers so some large movie company could put one video card in each
container for render farms. There are lots of devices that aren't
_really_ privileged (/dev/zero, /dev/null, /dev/full, /dev/ram) and
more that could be emulated...
(Really this seems like a situation where devtmpfs needs -o newinstance
and a corresponding filtered event netlink feed to pass to udev/mdev.
And that might even have been feasible before Greg decided that
devtmpfs needed to have opinions about uids. Oh well.)
More information about the Containers