[lxc-devel] [PATCH] Support MS_SHARED / - issues calling MAKEDEV

[Serge Hallyn]

Quoting Michael H. Warfield (mhw at WittsEnd.com):
> On Wed, 2013-01-09 at 21:27 +0800, Alexander Vladimirov wrote:
> > 2013/1/9 Michael H. Warfield <mhw at wittsend.com>:
> > > On Tue, 2013-01-08 at 15:40 -0500, Michael H. Warfield wrote:
> > >> On Wed, 2013-01-09 at 04:31 +0800, Alexander Vladimirov wrote:
> > >> > Well, properly placed hook could do all the things MAKEDEV supposed to do.
> > >>
> > >> And would give us the flexibility to worry about things like serial
> > >> devices or some USB devices or a few other things that would ordinarily
> > >> be populated by udev but don't vary over the life of the container.
> > >>
> > >> > 2013/1/9 Serge Hallyn <serge.hallyn at canonical.com>:
> > >> > > Quoting Michael H. Warfield (mhw at WittsEnd.com):
> > >> > >> More on the MAKEDEV debacle...
> > >> > > ...
> > >> > >> This whole thing with MAKEDEV is looking more and more like a morass
> > >> > >> with no way to cleanly resolve it.
> > >> > >
> > >> > > It sounds like consensus is it should be dropped from staging?
> > >
> > >> +1 for dropping the call to MAKEDEV, yeah.
> > >
> > > Looking at the sources now, removing run_makedev() and the call to is is
> > > pretty trivial.  I've already got a patch for that.
> > >
> > > WRT Alexander's suggestion for a hook...  I like that idea but a
> > > question comes up.  I've already got a patch for that hook,
> > > lxc.hook.autodev, as well and it's very close to where the
> > > lxc.hook.mount hooks is located only this one only gets invoked if
> > > autodev=1.  I'm still doing some testing but it looks like this could be
> > > done with the mount hook if we don't care for an isolated autodev one.
> > > OTOH, I can see some value in having a separate hook that only gets
> > > called if autodev is enabled.
> > >
> > > Next thought...  I dawns on me that many of these scripts could use some
> > > environment variables, such as the container name, the location of the
> > > rootfs, the location of the conf file, etc, etc.  That way, you could
> > > make the scripts a little more generic.  Problem is that we clear the
> > > environment and set "container=lxc" very early on in the process of
> > > starting up lxc-start.  Wouldn't that be just as effective if both where
> > > done just before execing the container and giving us the ability to pass
> > > environment variables to the hook scripts?  Maybe just AFTER
> > > lxc.hook.start?
> 
> > Maybe we could just provide some environment configuration option in
> > container config to extend basic env with required values?
> 
> That's an interesting thought too but probably only applicable to
> application containers.  I had thought about that as well.  Right now,
> my thought is to allow internally generated environment variables to be
> passed only to the hook scripts and be flushed before calling the
> container itself.
> 
> Moving the clearenv and putenv calls to just after the pre-start hook
> would do that plus allow for passing of environment variables from the
> invoking process through to the hook scripts (may be useful) but still
> flushed before invoking the container proper.
> 
> The reason I suggest this is that I now realize my earlier attempts at
> using the hook scripts to auto populate the container /dev all failed
> because I failed to realize WHERE the rootfs was mounted.  I thought I
> needed to access the directory in the path I had specified in the rootfs
> config variable when, in fact, I needed to use /usr/lib64/lxc/rootfs
> instead.  How is a hook script writer expected to know that without
> reading the sources and playing with the code?  That's a problem.

Also sounds like the manpages could stand to have some more info.

> But that's orthogonal with adding environment variables to be passed to
> the container itself.  Right now, the only one of any significance for
> the full system containers is "container=lxc" which we have hard coded
> and init (systemd init at least) flushes its environment so it won't
> propagate further.  Application containers may be a different matter I
> haven't explored.
> 
> It's a simple enough patch just moving where we clear and set the
> environment.  I may complete a whole combined patch that removes the
> MAKEDEV, adds the autodev hook, and moves the clear and put env calls
> shortly and throw it out for comments and attract rocks.  :-)=)

Great, thanks, that sounds good.

-serge