Re: clone3: allow creation of time namespace with offset

From: Christian Brauner
Date: Wed Apr 01 2020 - 07:46:35 EST


On Wed, Apr 01, 2020 at 01:40:25PM +0200, Michael Kerrisk (man-pages) wrote:
> On 3/25/20 12:26 PM, Christian Brauner wrote:
> > On Wed, Mar 25, 2020 at 08:58:36AM +0100, Adrian Reber wrote:
> >> On Tue, Mar 24, 2020 at 06:56:49PM +0100, Christian Brauner wrote:
> >>> On Tue, Mar 24, 2020 at 05:25:46PM +0100, Adrian Reber wrote:
> >>>> On Tue, Mar 24, 2020 at 05:09:45PM +0100, Christian Brauner wrote:
> >>>>> On Fri, Mar 20, 2020 at 11:33:55AM -0700, Andrei Vagin wrote:
> >>>>>> On Thu, Mar 19, 2020 at 11:29:55AM +0100, Christian Brauner wrote:
> >>>>>>> On Thu, Mar 19, 2020 at 09:16:43AM +0100, Arnd Bergmann wrote:
> >>>>>>>> On Thu, Mar 19, 2020 at 9:11 AM Adrian Reber <areber@xxxxxxxxxx> wrote:
> >>>>>>>>
> >>>>>>>>> With Arnd's idea of only using nanoseconds, timens_offset would then
> >>>>>>>>> contain something like this:
> >>>>>>>>>
> >>>>>>>>> struct timens_offset {
> >>>>>>>>> __aligned_s64 monotonic_offset_ns;
> >>>>>>>>> __aligned_s64 boottime_offset_ns;
> >>>>>>>>> };
> >>>>>>>>>
> >>>>>>>>> I kind of prefer adding boottime and monotonic directly to struct clone_args
> >>>>>>>>>
> >>>>>>>>> __aligned_u64 tls;
> >>>>>>>>> __aligned_u64 set_tid;
> >>>>>>>>> __aligned_u64 set_tid_size;
> >>>>>>>>> + __aligned_s64 monotonic_offset_ns;
> >>>>>>>>> + __aligned_s64 boottime_offset_ns;
> >>>>>>>>> };
> >>>>>>>>
> >>>>>>>> I would also prefer the second approach using two 64-bit integers
> >>>>>>>> instead of a pointer, as it keeps the interface simpler to implement
> >>>>>>>> and simpler to interpret by other tools.
> >>>>>>>
> >>>>>>> Why I don't like has two reasons. There's the scenario where we have
> >>>>>>> added new extensions after the new boottime member and then we introduce
> >>>>>>> another offset. Then you'd be looking at:
> >>>>>>>
> >>>>>>> __aligned_u64 tls;
> >>>>>>> __aligned_u64 set_tid;
> >>>>>>> __aligned_u64 set_tid_size;
> >>>>>>> + __aligned_s64 monotonic_offset_ns;
> >>>>>>> + __aligned_s64 boottime_offset_ns;
> >>>>>>> __aligned_s64 something_1
> >>>>>>> __aligned_s64 anything_2
> >>>>>>> + __aligned_s64 sometime_offset_ns
> >>>>>>>
> >>>>>>> which bothers me just by looking at it. That's in addition to adding two
> >>>>>>> new members to the struct when most people will never set CLONE_NEWTIME.
> >>>>>>> We'll also likely have more features in the future that will want to
> >>>>>>> pass down more info than we want to directly expose in struct
> >>>>>>> clone_args, e.g. for a long time I have been thinking about adding a
> >>>>>>> struct for CLONE_NEWUSER that allows you to specify the id mappings you
> >>>>>>> want the new user namespace to get. We surely don't want to force all
> >>>>>>> new info into the uppermost struct. So I'm not convinced we should here.
> >>>>>>
> >>>>>> I think here we can start thinking about a netlink-like interface.
> >>>>>
> >>>>> I think netlink is just not a great model for an API and I would not
> >>>>> want us to go down that route.
> >>>>>
> >>>>> I kept thinking about this for a bit and I think that we will end up
> >>>>> growing more namespace-related functionality. So one thing that came to
> >>>>> my mind is the following layout:
> >>>>>
> >>>>> struct {
> >>>>> struct {
> >>>>> __s64 monotonic;
> >>>>> __s64 boot;
> >>>>> } time;
> >>>>> } namespaces;
> >>>>>
> >>>>> struct _clone_args {
> >>>>> __aligned_u64 flags;
> >>>>> __aligned_u64 pidfd;
> >>>>> __aligned_u64 child_tid;
> >>>>> __aligned_u64 parent_tid;
> >>>>> __aligned_u64 exit_signal;
> >>>>> __aligned_u64 stack;
> >>>>> __aligned_u64 stack_size;
> >>>>> __aligned_u64 tls;
> >>>>> __aligned_u64 set_tid;
> >>>>> __aligned_u64 set_tid_size;
> >>>>> __aligned_u64 namespaces;
> >>>>> __aligned_u64 namespaces_size;
> >>>>> };
> >>>>>
> >>>>> Then when we end up adding id mapping support for CLONE_NEWUSER we can
> >>>>> extend this with:
> >>>>>
> >>>>> struct {
> >>>>> struct {
> >>>>> __aligned_u64 monotonic;
> >>>>> __aligned_u64 boot;
> >>>
> >>> s/__aligned_u64/__s64/g
> >>>
> >>> Sorry, leftover from my first draft.
> >>>
> >>>>> } time;
> >>>>>
> >>>>> struct {
> >>>>> /* id mapping members */
> >>>>> } user;
> >>>>> } namespaces;
> >>>>>
> >>>>> Thoughts? Other ideas?
> >>>>
> >>>> Works for me.
> >>>>
> >>>> If we add the user namespace id mappings and then at some point a third
> >>>> element for the time namespace appears it would also start to be mixed.
> >>>> Just as you mentioned that a few mails ago.
> >>>
> >>> I think you misunderstand me or I'm misunderstanding you. That new time
> >>> namespace member would go into struct time {} so
> >>>
> >>> struct {
> >>> struct {
> >>> __s64 monotonic;
> >>> __s64 boot;
> >>> __s64 someothertime;
> >>> } time;
> >>>
> >>> struct {
> >>> /* id mapping members */
> >>> } user;
> >>> } namespaces;
>
> So far, this seems like the least worst approach to me :-).
>
> I think it's reasonable to assume that there will be another
> time NS offset to add one day. I don't think anyone expected
> CLOCK_BOOTIME (added in 2011) at the time that CLOCK_MONOTONIC
> appeared (as part of the POSIX timers API in Linux 2.6.0 2003);
> similarly, we probably can't conceive now what clock might be
> added in the future that should also be governed by time
> namespaces.
>
>
> But...
>
> >> My question was about how does the kernel know how 'struct namespaces'
> >> is structured. How can an older kernel (which only is aware of two
> >> clocks) deal with a, like in this example, third clock. Will the size
> >> '__aligned_u64 namespaces_size' be used for versioning?
> >
> > Yes, that would be the idea.
>
> The idea implied here (if I understand correctly) of "binary chop
> on the structure size to see what your kernel supports" is pretty
> clunky, IMO. It's worth at least considering alternatives.
> For example, seccomp has a number of interesting interfaces to
> discover what the running kernel supports [1]. Maybe it is worth
> considering something similar for clone3()?

Yes, that's definitely something I've been thinking about and I've also
discussed this publicly on the libc-alpha mailing list since Florian
(Weimer) wants this for glibc. _But_ I didn't want to simply go my own
way and so Aleksa and I stuck our heads together and tried to come up
with a generic way that e.g. would work for both openat2() and for
clone3(), in fact we tried to come up with a generic way that could
potentially be used by any syscall that makes use of the new
copy_struct_from_user() helper that Aleksa and I pushed upstream. We
haven't yet gotten around to actually try and implementing our ideas.
I'll try to get around to this during this cycle (/me makes note to "sit
down" with Aleksa). We'll definitely take a look at seccomp and cgroups
too.

Christian

>
> Cheers,
>
> Michael
>
> [1]
> /proc/sys/kernel/seccomp/actions_avail (see seccomp(2))
> SECCOMP_GET_ACTION_AVAIL (see secommp(2))
> cgroups also provides something similar in the form of
> /sys/kernel/cgroup/features (see cgroups(7))
>
> --
> Michael Kerrisk
> Linux man-pages maintainer; http://www.kernel.org/doc/man-pages/
> Linux/UNIX System Programming Training: http://man7.org/training/