Re: [libvirt] [RFC PATCH 2/2] LXC: Create ro overlay mounts only if we're not within a user namespace

Am 01.07.2013 13:35, schrieb Daniel P. Berrange: > On Mon, Jul 01, 2013 at 01:25:28PM +0200, Richard Weinberger wrote: >> Am 01.07.2013 13:22, schrieb Daniel P. Berrange: >>> On Mon, Jul 01, 2013 at 01:05:23PM +0200, Richard Weinberger wrote: >>>> Am 01.07.2013 12:33, schrieb Daniel P. Berrange: >>>>> On Mon, Jul 01, 2013 at 08:29:14AM +0200, Richard Weinberger wrote: >>>>>> Any ideas what's going on here? >>>>> >>>>> No, it is very odd. It smells like a kernel issue to me. What >>>>> version are you running ? >>>> >>>> I see this issue on all kernels. >>>> Currently I'm using vanilla v3.9.x and v3.10. >>>> >>>>> I've also tried running the demo programs shown on the LWN.net >>>>> article >>>>> >>>>> https://lwn.net/Articles/532593/ >>>>> >>>>> and they don't operate in the way described by the article - the demo >>>>> programs continue to ru as 'nfsnobody' even after the mappings are >>>>> setup. >>>>> >>>>> I'm just using the Fedora 3.9.4-303 kernel, rebuilt with userns enabled >>>>> in KConfig. I'm wondering if there is still stuff missing in 3.9.x >>>>> that prevents this from working properly, or if the kernel behaviour >>>>> changed after those LWN articles were written. >>>> >>>> To me it looks like the capability system behaves odd. >>>> The mappings in /proc are fine as long I do not call capng_updatev(). >>>> Also calling capng_updatev() with parameters that do not change the current cap set >>>> triggers the odd behavior too. >>>> >>>> So we see two (related?) issues: >>>> 1. If we try updating the capabilities of pid1 /proc/1/ has unmapped files till we exec(). >>>> 2. Dropping capabilities does not work we always gain a fresh and full capability set. >>>> >>>> BTW: I'm sure the issues are not caused by Gau Feng's userns patches. >>> >>> Yeah, I've reproduced this problem with standalone code outside of >>> libvirt. >>> >>> Take the attached code and run >> >> -ENOATTACHMENT :-( > > Now really attached. > > I think I might know what is happening now though. When you start a new > namespace, you must mount a new instance of 'proc' filesystem. We are > not synchronizing this wrt setup of the uid/gid mappings though, so we > are racy. So I have a feeling we're creating the proc filesystem before > the mappings are setup. I'm going to add some synchronization in to see > if it makes a difference in this respect. So you mount /proc and write the uid/gid mappings in parallel? Both has to be done on the host side. Why is this parallel?