Re: Time Namespaces: CLONE_NEWTIME and clone3()?

From: Adrian Reber
Date: Tue Feb 18 2020 - 12:20:27 EST


On Tue, Feb 18, 2020 at 12:03:31AM +0100, Christian Brauner wrote:
> On Mon, Feb 17, 2020 at 10:47:53PM +0100, Michael Kerrisk (man-pages) wrote:
> > Hello Christian,
> >
> > On Mon, 17 Feb 2020 at 16:15, Christian Brauner
> > <christian.brauner@xxxxxxxxxx> wrote:
> > >
> > > On Mon, Feb 17, 2020 at 03:20:55PM +0100, Michael Kerrisk wrote:
> > > > Hello Dmitry, Andrei,
> > > >
> > > > Is the CLONE_NEWTIME flag intended to be usable with clone3()? The
> > > > mail quoted below implies (in my reading) that this should be possible
> > > > once clone3() is available, which it is by now. (See also [1].)
> > > >
> > > > If the answer is yes, CLONE_NEWTIME should be usable with clone3(),
> > > > then I have a bug report and a question.
> > > >
> > > > I successfully used CLONE_NEWTIME with unshare(). But if I try to use
> > > > CLONE_NEWSIGNAL with clone3(), it errors out with EINVAL, because of
> > >
> > > s/CLONE_NEWSIGNAL/CLONE_NEWTIME/
> > >
> > > > the following check in clone3_args_valid():
> > > >
> > > > /*
> > > > * - make the CLONE_DETACHED bit reuseable for clone3
> > > > * - make the CSIGNAL bits reuseable for clone3
> > > > */
> > > > if (kargs->flags & (CLONE_DETACHED | CSIGNAL))
> > > > return false;
> > > >
> > > > The problem is that CLONE_NEWTIME matches one of the bits in the
> > > > CSIGNAL mask. If the intention is to allow CLONE_NEWTIME with
> > > > clone3(), then either the bit needs to be redefined, or the error
> > > > checking in clone3_args_valid() needs to be reworked.
> > >
> > > If this is intended to be useable with clone3() the check should be
> > > adapted to allow for CLONE_NEWTIME. (I asked about this a while ago I
> > > think.)
> > > But below rather sounds like it should simply be an unshare() flag. The
> > > code seems to set frozen_offsets to true right after copy_namespaces()
> > > in timens_on_fork(new_ns, tsk) and so the offsets can't be changed
> > > anymore unless I'm reading this wrong.
> > > Alternatives seem to either make timens_offsets writable once after fork
> > > and before exec, I guess - though that's probably not going to work
> > > with the vdso judging from timens_on_fork().
> > >
> > > The other alternative is that Andrei and Dmitry send me a patch to
> > > enable CLONE_NEWTIME with clone3() by exposing struct timens_offsets (or
> > > a version of it) in the uapi and extend struct clone_args to include a
> > > pointer to a struct timens_offset that is _only_ set when CLONE_NEWTIME
> > > is set.
> > > Though the unshare() way sounds way less invasive simpler.
> >
> > Actually, I think the alternative you propose just here is better. I
> > imagine there are times when one will want to create multiple
> > namespaces with a single call to clone3(), including a time namespace.
> > I think this should be allowed by the API. And, otherwise, clone3()
> > becomes something of a second-class citizen for creating namespaces.
> > (I don't really get the "less invasive" argument. Implementing this is
> > just a piece of kernel to code to make user-space's life a bit simpler
> > and more consistent.)
>
> I don't particularly mind either way. If there's actual users that need
> to set it at clone3() time then we can extend it. So I'd like to hear
> what Adrian, Dmitry, and Thomas think since they are well-versed how
> this will be used in the wild. I'm weary of exposing a whole new uapi
> struct and extending clone3() without any real use-case but I'm happy to
> if there is!

Re-creating a time namespace during restore via clone3() would be CRIU's
preferred way of doing this. If available CRIU is already using
clone3(), especially because of set_tid, and for the time namespace we
would also rely on clone3() to re-create it.

I will provide a patch to extend clone3() to handle the creation of a
new time namespace.

Adrian