On Sat, 26 Nov 2016 at 10:27:38 +0000, Ben Hutchings wrote:
> On Fri, 2016-11-25 at 18:54 +0100, Marco d'Itri wrote:
> > setup_devices_simple() could be easily modified to create /dev/ptmx with
> > mknod instead of the current symlink.
That's presumably:
- ln -s pts/ptmx $TARGET/dev/ptmx
+ mknod -m 666 $TARGET/dev/ptmx c 5 2
That matches what cdebootstrap does, for what it's worth. (Was
cdebootstrap meant to be used in d-i like cdebconf is, or is it just
a trap for the unwary?)
> > While this would solve this problem, it would also break again
> > debootstrap in containers (which was discussed in different bugs), so
> > I am not sure if it is a good idea.
Am I correct in thinking that the issue you're talking about here is that
the obvious solution (above) would break this scenario?
* Obtain a rootfs for a container somehow (debootstrap or otherwise)
* Launch a container with that rootfs, using some containerization
technology (which one?) that applies the same device-cgroup policy that
systemd's src/core/cgroup.c does:
- may read and write null, zero, full, random, urandom, tty, ptmx
- may mknod null, zero, full, random, urandom, tty but *not* ptmx
* Inside that container, run debootstrap
* Good result: debootstrap succeeds
* Bad result: debootstrap fails to mknod $TARGET/dev/ptmx
Or if it's something different, please could you describe it in similar
terms?
If debootstrap inside a container is meant to work, then this would seem
like a job for an autopkgtest. I'll try to write one if someone can tell me
which containerization technologies (systemd-nspawn? lxc? both? others?)
are the ones of interest here.
However, it seems to me that being able to create a chroot that can be
chrooted into with currently-critical bits of Debian infrastructure
(schroot, sbuild) *at all* is probably a higher priority for stretch
than being able to create a conditionally-usable chroot from inside a
container. (By "conditionally usable" I mean schroot/pbuilder can't use
the resulting chroot but maybe systemd-nspawn can.)
Regardless of whether people think sbuild+schroot is the best thing for
the future, it's how current Debian buildds work, and how the next Debian
stable's buildds will work unless sbuild changes extensively during the
freeze, so anything that prevents it from working is de facto release
critical for all releases <= buster.
> The temporary workaround with /dev/ptmx could be made optional. It's
> not OK to break the previously working configurations.
If I'm understanding the situation correctly then the next best thing would
seem to be:
- ln -s pts/ptmx $TARGET/dev/ptmx
+ # Inside a container, we might not be allowed to create /dev/ptmx.
+ # If not, do the next best thing (but see #817236).
+ mknod -m 666 $TARGET/dev/ptmx c 5 2 || ln -s pts/ptmx $TARGET/dev/ptmx
which would result in debootstrap inside a container continuing to create
a /dev that current schroot etc. cannot successfully use (but that's maybe
better than it failing completely?), whereas debootstrap outside a container
would create a /dev that works fine?
Is there a reason why mounting /dev/pts results in 000 permissions
on /dev/pts/ptmx? That seems odd. If it didn't, then what debootstrap
does would work.
I notice that systemd creates a symlink when making a new namespace, but
systemd also mounts /dev/pts with newinstance,ptmxmode=0666 when making a
new namespace, and existing tools like schroot and pbuilder presumably
don't do that. Should they? Or would that break the ability for an
interactive shell inside the chroot to have its stdin/stdout/stderr
point to the pty created by an xterm, screen or equivalent outside the
chroot?
Regards,
S
